Comparative Analysis of Embryo Ploidy Prediction Models: From Traditional PGT-A to Advanced AI Algorithms

Abstract

This comprehensive review analyzes the evolving landscape of embryo ploidy prediction models, comparing traditional invasive methods like PGT-A with emerging non-invasive artificial intelligence approaches. We examine foundational principles of aneuploidy detection, methodological advances in machine learning and deep learning applications, optimization challenges across diverse clinical settings, and validation metrics for model performance. For researchers and drug development professionals, this synthesis provides critical insights into how AI-driven tools like BELA, iDAScore, and morphokinetic algorithms are transforming embryo selection paradigms while highlighting persistent limitations and future research directions in reproductive medicine.

The Evolution of Ploidy Assessment: From FISH to Non-Invasive Paradigms

Chromosomal aneuploidy, defined as an abnormal number of chromosomes, represents a major genetic disorder with profound implications for human reproduction, embryonic development, and cancer biology. This condition is a leading cause of infertility, pregnancy loss, and developmental disabilities, with over 25% of all miscarriages being monosomic or trisomic [1]. Aneuploidy is present in an estimated 10-30% of all fertilized eggs, establishing it as a critical factor in human reproduction and development [1].

The clinical significance of aneuploidy spans multiple medical disciplines, from reproductive medicine to oncology. In prenatal genetics, aneuploidies of chromosomes 13, 18, and 21 result in Patau, Edwards, and Down syndromes, respectively—the only full autosomal trisomies compatible with postnatal survival [1]. Meanwhile, in oncology, aneuploidy has been cemented as a hallmark of cancer, with recent research revealing the complex relationship between specific chromosomal alterations and tumor behavior [2].

This comparative analysis examines the prevalence, clinical impact, and correlation with maternal age of chromosomal aneuploidy, with a specific focus on evaluating emerging technologies for its detection and prediction. The review synthesizes current evidence across diverse clinical contexts, from preimplantation genetic testing to prenatal screening and cancer research, providing researchers with a comprehensive framework for understanding this biologically and clinically significant phenomenon.

Prevalence and Clinical Impact Across Populations

Spectrum of Chromosomal Aneuploidies

Aneuploidy manifestations vary considerably across clinical contexts, with specific chromosomal abnormalities demonstrating distinct prevalence patterns and clinical outcomes. Table 1 summarizes the prevalence rates and clinical correlates of major aneuploidy types across different populations.

Table 1: Prevalence and Clinical Correlates of Major Aneuploidy Types

Aneuploidy Category	Specific Type	Prevalence/Detection Rate	Key Clinical Correlates	Population Context
Embryonic Aneuploidy	Overall prevalence	10-30% of fertilized eggs [1]	Leading cause of implantation failure and pregnancy loss [1]	Preimplantation embryos
	Products of conception	67.8% of spontaneous pregnancy losses [3]	Chromosomes 7 and 16 most commonly affected [3]	First-trimester pregnancy loss
Autosomal Trisomies	Trisomy 21 (Down syndrome)	Most common viable autosomal trisomy [1]	Characteristic physical features, neurocognitive impairment [4]	Live births
	Trisomy 18 (Edwards syndrome)	Second most common viable autosomal trisomy [1]	Multiple congenital anomalies, reduced survival [1]	Live births
	Trisomy 13 (Patau syndrome)	Third most common viable autosomal trisomy [1]	Severe structural defects, profound developmental disability [1]	Live births
Sex Chromosome Aneuploidies (SCAs)	Overall prevalence	~1 in 440 newborns [5]	Variable phenotype; may include infertility, learning difficulties [5]	General population
	Turner syndrome (45,X)	PPV of NIPT: 27.8% [5]	More common in adolescent pregnancies [6]	Prenatal screening
	Klinefelter syndrome (47,XXY)	PPV of NIPT: 100% [5]		Prenatal screening
	Triple X syndrome (47,XXX)	PPV of NIPT: 50.0% [5]		Prenatal screening
	Jacobs syndrome (47,XYY)	PPV of NIPT: 100% [5]		Prenatal screening
Rare Chromosomal Abnormalities (RCAs)	Overall detection	0.36% in NIPT screening [7]	Low PPV (6.86%); associated with adverse pregnancy outcomes [7]	General obstetric population
	Trisomy 7	Most prevalent RCA (27.98% of RCAs) [7]	Occurs independently of maternal age [7]	NIPT screening

Aneuploidy in Pregnancy Loss

Comprehensive genomic analysis of first-trimester spontaneous pregnancy losses has revealed that approximately 67.8% contain chromosomal abnormalities, a higher percentage than previously reported in studies using conventional karyotyping alone [3]. This finding emerges from advanced techniques including genome haplarithmisis, which detects aberrations missed by traditional cytogenetic methods.

The distribution of abnormal cells varies between embryonic and placental lineages in spontaneous pregnancy losses. Contrary to the pattern observed in viable pregnancies—where mosaic chromosomal abnormalities are often restricted to chorionic villi (confined placental mosaicism)—research demonstrates a higher degree of mosaic chromosomal imbalances in extra-embryonic mesoderm compared to chorionic villi in pregnancy losses [3]. This reversed distribution pattern suggests fundamental differences in how developing systems manage chromosomal abnormalities in successful versus failed pregnancies.

Aneuploidy in Adolescent and Advanced Maternal Age Pregnancies

Maternal age exerts a profound influence on aneuploidy risk, with both extremes of the reproductive age spectrum associated with elevated rates of chromosomal abnormalities. Adolescent pregnancies demonstrate a unique profile of chromosomal abnormalities characterized by:

• Lower overall trisomy rates (5.9%) compared to women aged 20-34 (9.3%) and ≥35 years (12.1%) [6]
• Higher prevalence of Turner syndrome (4.6%) compared to women 20-34 (2.8%) or ≥35 years (0.1%) [6]
• Increased risk of unspecified fetal sex (RR = 2.25) and culture failure (RR = 4.32) in prenatal diagnostic procedures [6]

Advanced maternal age (≥35 years) is associated with well-documented risks including spontaneous abortion, infertility, and genetic disorders in offspring [4]. The predominant mechanism underlying age-related aneuploidy involves meiotic errors in oocytes, particularly during the first meiotic division [4]. Molecular studies indicate that premature separation or reverse segregation of sister chromatids is more prevalent in aged oocytes, whereas nondisjunction underlies aneuploidy in adolescent conceptions [4].

Comparative Analysis of Aneuploidy Detection and Prediction Technologies

Performance Metrics of Detection Methodologies

Table 2 provides a comprehensive comparison of current technologies for aneuploidy detection and prediction, highlighting their performance characteristics, advantages, and limitations.

Table 2: Comparative Performance of Aneuploidy Detection and Prediction Technologies

Technology	Application Context	Performance Metrics	Advantages	Limitations
Karyotyping	Prenatal diagnosis (gold standard)	High resolution for full chromosome analysis	Comprehensive chromosome analysis, detects balanced rearrangements	Time-consuming (2-3 weeks), requires cell culture [1]
QF-PCR	Rapid aneuploidy detection	~48 hours for 96 samples [1]	Rapid, automated, cost-effective	Limited by genetic polymorphism variability [1]
AI-enhanced QF-PCR	Rapid aneuploidy detection	Accuracy: High; Analysis time: 1.7 seconds (vs. 45 min manual) [1]	Dramatically reduced analysis time, minimized human error	Requires technical validation across populations
NIPT for common trisomies	Prenatal screening	High accuracy for trisomies 21, 18, 13 [5]	Non-invasive, high sensitivity and specificity	Screening, not diagnostic
NIPT for SCAs	Prenatal screening	Variable PPV: 27.8%-100% depending on SCA type [5]	Non-invasive, detects sex chromosome abnormalities	Lower specificity than for autosomal trisomies
NIPT for RCAs	Prenatal screening	Low PPV (6.86%) [7]	Broad screening capability	High false positive rate, challenging counseling
PGT-A	Embryo selection in IVF	Gold standard for embryonic aneuploidy [8]	Direct assessment of embryonic chromosomes	Invasive, costly, not universally accessible [8]
iDAScore v1.0	Embryo ploidy prediction	AUC: 0.60-0.67 for euploidy prediction [8]	Non-invasive, utilizes time-lapse imaging	Moderate predictive accuracy
iDAScore v2.0	Embryo ploidy prediction	AUC: 0.635-0.68 for euploidy prediction [8]	Improved performance over v1.0	Cannot replace PGT-A
FEMI Foundation Model	Embryo ploidy prediction	AUROC >0.75 using only image data [9]	Self-supervised learning on 18 million images, multiple task capabilities	Requires diverse training data for optimal performance

Deep Learning Models for Embryo Ploidy Prediction

Artificial intelligence-based embryo selection tools represent a promising non-invasive approach for evaluating embryo viability and ploidy status in in vitro fertilization (IVF). Among these, iDAScore has emerged as a well-validated deep learning model that analyzes time-lapse embryo images to assign scores reflecting the likelihood of implantation and live birth [8].

Multiple retrospective studies have demonstrated a statistically significant association between higher iDAScore values and embryo euploidy, with AUC values for euploidy prediction ranging from 0.60 to 0.68 across studies [8]. The predictive performance shows modest improvement when iDAScore is combined with clinical and embryonic parameters (AUC increasing to 0.688), suggesting a complementary role alongside traditional parameters rather than replacement of established methods [8].

The recently developed FEMI (Foundational IVF Model for Imaging) foundation model represents a significant advancement in the field, having been trained on approximately 18 million time-lapse embryo images using a Vision Transformer masked autoencoder (ViT MAE) architecture [9]. This model achieves an AUROC >0.75 for ploidy prediction using only image data—significantly outpacing benchmark models [9]. FEMI's architecture enables multiple downstream tasks including blastocyst quality scoring, embryo component segmentation, and developmental milestone timing, demonstrating the potential of foundation models to standardize and improve embryo assessment in IVF.

Experimental Protocols for Key Technologies

AI-Enhanced QF-PCR Protocol

The AI-driven quantitative fluorescent polymerase chain reaction (QF-PCR) approach introduces significant innovations to traditional aneuploidy detection:

• Sample Processing: DNA extraction from amniotic fluid samples using QIAamp DNA Mini Kit, with quality assessment via NanoDrop spectrophotometer [1]
• PCR Amplification: Targeting segmental duplications on chromosomes 13, 18, 21, X, and Y using fluorescently labeled primers [1]
• Fragment Analysis: Capillary electrophoresis on a 3500 Genetic Analyzer with GeneMapper software [1]
• AI Integration: Fluorescence intensity data processed through a Python-based computational pipeline implementing an XGBoost classifier trained on 80% of the dataset and tested on the remaining 20% [1]
• Validation: Results confirmed against conventional karyotyping as the gold standard [1]

This integrated approach reduces analysis time from 45 minutes (manual interpretation) to 1.7 seconds while minimizing human errors, demonstrating the transformative potential of AI in diagnostic laboratory workflows [1].

FEMI Foundation Model Training Protocol

The development of the FEMI foundation model involved a sophisticated multi-stage process:

• Data Collection: Compilation of 17,968,959 time-lapse images from multiple clinics and public datasets [9]
• Preprocessing: Image cropping around embryos using a segmentation model based on InceptionV3 architecture, followed by resizing to 224×224 pixels [9]
• Model Architecture: Vision Transformer masked autoencoder (ViT MAE) backbone pre-trained on ImageNet-1k and further pre-trained on the embryo image dataset for 800 epochs with early stopping [9]
• Self-Supervised Learning: Training through image reconstruction from masked inputs to learn domain-specific features [9]
• Downstream Task Adaptation: Application to ploidy prediction, blastocyst quality scoring, embryo component segmentation, embryo witnessing, blastulation time prediction, and stage prediction through task-specific layers appended to the encoder [9]

This protocol demonstrates how self-supervised learning on large-scale, unlabeled datasets can produce models with robust performance across multiple clinically relevant tasks in embryology.

Biological Mechanisms of Aneuploidy

The development of aneuploidy involves complex biological processes operating at multiple levels. The following diagram illustrates key molecular mechanisms contributing to age-related aneuploidy in oocytes:

Diagram 1: Molecular mechanisms of age-related aneuploidy in oocytes. Key pathways through which advanced maternal age contributes to meiotic errors and chromosomal abnormalities in oocytes, highlighting structural, genomic, and cellular processes. SAC = Spindle Assembly Checkpoint.

The biological mechanisms underlying aneuploidy formation vary significantly across maternal age groups and clinical contexts. In adolescent pregnancies, the predominant mechanism involves nondisjunction events during meiosis, whereas in advanced maternal age, premature separation or reverse segregation of sister chromatids represents the more common mechanism [4]. These differences reflect distinct biological vulnerabilities across the reproductive lifespan.

Multiple molecular pathways contribute to age-related aneuploidy, with cohesin complex weakening and weakened spindle assembly checkpoint (SAC) signaling identified as key factors [4]. Cohesin complexes, comprised of kleisin, SMC1/3, and STAG subunits, are integral to meiotic chromosome dynamics, and their age-related deterioration contributes significantly to improper chromosome segregation [4]. Simultaneously, genomic instability mechanisms including accumulated DNA damage, epigenetic dysregulation, and mitochondrial decline further drive meiotic abnormalities in aging oocytes [4].

Research Reagent Solutions for Aneuploidy Investigation

Table 3: Essential Research Reagents for Aneuploidy Investigation

Reagent/Kit	Application	Key Features	Representative Use
QIAamp DNA Mini Kit	DNA extraction from clinical samples	Efficient nucleic acid purification	DNA extraction from amniotic fluid for QF-PCR [1]
Ion Plus Fragment Library Kit	NIPT library preparation	End repair for sequencing libraries	Preparation of cfDNA libraries for NIPT [7]
CytoScanTM 750K	Chromosomal microarray analysis	High-resolution CNV detection	Prenatal diagnosis following abnormal screening [7]
EmbryoScope+/EmbryoScope	Time-lapse embryo imaging	Continuous embryo monitoring without disturbance	Image acquisition for iDAScore and FEMI analysis [8] [9]
3500 Genetic Analyzer	Fragment analysis	Capillary electrophoresis for size separation	QF-PCR product analysis [1]
GeneMapper Software	Fragment analysis data interpretation	Automated allele calling and sizing	Analysis of QF-PCR results [1]
ViT MAE Architecture	Foundation model training	Self-supervised learning for image analysis	FEMI model pre-training on embryo images [9]
XGBoost Classifier	Machine learning implementation	Gradient boosting framework for classification	AI-based analysis of QF-PCR fluorescence data [1]

Chromosomal aneuploidy represents a biologically complex and clinically significant challenge across multiple medical disciplines. The prevalence of approximately 67.8% in spontaneous pregnancy losses underscores its importance in reproductive failure, while its role as a hallmark of cancer highlights the diverse contexts in which chromosomal numerical abnormalities exert biological effects.

The correlation between maternal age and aneuploidy risk demonstrates a U-shaped distribution, with both adolescent and advanced maternal age pregnancies showing elevated rates of specific chromosomal abnormalities, albeit through distinct biological mechanisms. This understanding enables more targeted counseling and management strategies for at-risk populations.

Emerging technologies, particularly AI-enhanced detection methods and deep learning models for embryo ploidy prediction, are revolutionizing the field of aneuploidy assessment. The development of foundation models like FEMI, trained on millions of time-lapse images, points toward a future with more standardized, objective, and comprehensive aneuploidy evaluation across clinical contexts. While current performance metrics of these technologies show promise, they generally serve as complementary tools rather than replacements for established diagnostic methods like PGT-A.

Future research directions should focus on elucidating the specific molecular pathways contributing to age-related aneuploidy, validating emerging AI models in diverse clinical populations, and developing targeted interventions to mitigate aneuploidy risk across the reproductive lifespan. The integration of multi-omics technologies with advanced computational approaches holds particular promise for advancing both fundamental understanding and clinical management of this complex biological phenomenon.

Preimplantation genetic testing for aneuploidy (PGT-A) has emerged as a pivotal technology in assisted reproductive technology (ART), providing a method for screening embryos for chromosomal abnormalities before uterine transfer. The procedure aims to select euploid embryos, thereby potentially improving implantation rates, reducing miscarriage risks, and shortening the time to pregnancy [10] [11]. Originally termed preimplantation genetic screening (PGS), the technology has evolved through several iterations, with current comprehensive chromosome screening technologies now referred to as PGT-A [12]. This review critically examines PGT-A's position as a contemporary gold standard, objectively comparing its performance against emerging alternatives through a detailed analysis of its technical procedures, analytical foundations, and documented limitations. The analysis is framed within the broader context of comparative embryo ploidy prediction models, providing researchers and scientists with a rigorous assessment of the current state of the art.

Biopsy Techniques for PGT-A

The biopsy process, which involves retrieving cellular material from oocytes or embryos, is a fundamental and technically demanding component of PGT-A. The method and timing of the biopsy significantly influence the reliability of the genetic diagnosis and the subsequent developmental potential of the embryo [10] [13].

Evolution of Biopsy Methods

The technique for accessing embryonic cells has evolved from mechanical and chemical opening of the zona pellucida to the current laser-assisted approach. According to data from the ESHRE PGT Consortium, by 2015, the laser method was employed in 98% of PGT procedures, largely replacing earlier methods due to being less operator-dependent, having a shorter learning curve, and causing no alterations to outcomes [10].

Types of Biopsy by Developmental Stage

The choice of biopsy stage represents a critical trade-off between embryo viability and diagnostic accuracy.

• Polar Body (PB) Biopsy: This method involves the removal of the first and second polar bodies from the oocyte or day-1 embryo. While minimally invasive, its significant limitation is that it analyses exclusively maternal genetic material, providing no information on the paternal genetic contribution or post-fertilization mitotic errors [10] [13]. Consequently, its clinical use is now limited, accounting for just 1% of PGT cases in 2018, and it is primarily utilized in countries with legal restrictions on embryo biopsy [10] [13].

• Blastomere Biopsy (Cleavage Stage): Performed on day-3 embryos, this technique involves the extraction of one blastomere from a 6-8 cell embryo. Its main advantage is the ability to perform a fresh transfer. However, the analysis of only 1-2 cells presents substantial limitations, including technical challenges such as allele drop-out, preferential amplification, and a high rate of DNA amplification failure, which can lead to misdiagnosis [10]. Furthermore, the removal of a single blastomere has been shown to negatively affect subsequent embryo development, including delayed compaction and impaired hatching [10]. Its use has declined sharply, particularly for PGT-A, dropping from 8% in 2016-2017 to just 0.6% in 2018 [10].

• Trophectoderm (TE) Biopsy (Blastocyst Stage): This is the current gold-standard method. Performed on day 5/6 embryos, it involves the extraction of 5-10 cells from the trophectoderm, which is the precursor to the placenta [10] [14] [13]. This method offers several key advantages: a larger amount of DNA for analysis, reducing inconclusive diagnoses to less than 5%; a lower impact on embryonic development as the cells biopsied are not part of the inner cell mass (the fetal precursor); and a better ability to detect mosaicism [10] [13]. Additionally, vitrified blastocysts have higher survival rates, facilitating deferred single embryo transfer (SET) and reducing the risk of multiple pregnancies [13].

Table 1: Comparison of Embryo Biopsy Techniques

Biopsy Type	Developmental Stage	Cells Retrieved	Advantages	Disadvantages
Polar Body (PB)	Oocyte / Day 1	1-2 (maternal)	Minimally invasive to embryo	Maternal genetics only; misses paternal errors & mitotic errors
Blastomere	Cleavage (Day 3)	1	Allows for fresh embryo transfer	High impact on viability; high risk of misdiagnosis; cannot detect mosaicism
Trophectoderm (TE)	Blastocyst (Day 5/6)	5-10	More DNA; less invasive; can detect mosaicism; higher diagnostic accuracy	Requires advanced blastocyst culture; not all embryos reach this stage

The following diagram illustrates the primary workflow for the trophectoderm biopsy, the current standard of care:

Analytical Principles and Platforms

The genetic analysis of biopsied cells has undergone a significant technological evolution, moving from limited chromosome screening to comprehensive 24-chromosome analysis.

Evolution of Testing Platforms

The initial iteration of PGT-A, often called PGS 1.0, used fluorescence in situ hybridization (FISH) to evaluate only 5-10 chromosomes. This method was later shown to have no beneficial effect on IVF outcomes [12]. The subsequent development of genome-wide platforms marked the beginning of PGS 2.0 and 3.0, utilizing technologies such as array comparative genomic hybridization (aCGH), single nucleotide polymorphism (SNP) arrays, quantitative polymerase chain reaction (qPCR), and next-generation sequencing (NGS) [15] [12]. NGS is currently considered the gold standard due to its superior efficiency, precision, and ability to detect mosaicism, all at a progressively lower cost [15] [12].

Diagnostic Workflow and Classification

Following the TE biopsy, the retrieved cells undergo whole-genome amplification (WGA) to generate sufficient DNA for analysis [13]. The DNA is then processed using the chosen platform (e.g., NGS) to determine the copy number of each chromosome. Embryos are subsequently classified into one of three categories:

• Euploid: Chromosomally normal embryos, which are prioritized for transfer.
• Aneuploid: Embryos with a uniform chromosomal abnormality, which are typically not transferred.
• Mosaic: Embryos containing a mixture of euploid and aneuploid cells. These represent a diagnostic challenge and are prioritized for transfer only after all euploid embryos have been used, and after extensive genetic counseling [14] [12].

Performance Data and Limitations of PGT-A

A critical assessment of PGT-A requires a clear-eyed examination of its diagnostic accuracy and clinical limitations, which are areas of active debate and research.

Diagnostic Accuracy and Misdiagnosis Rates

A recent comprehensive systematic review and meta-analysis (2025) provides robust quantitative data on the accuracy of PGT-A. The analysis, which included studies comparing TE biopsy results to a reference standard such as the whole dissected embryo/inner cell mass (WE/ICM) or prenatal/postnatal testing, found high predictive values for uniformly classified embryos [14] [16].

Table 2: Diagnostic Accuracy of PGT-A from Meta-Analysis (2025)

Embryo Classification	Predictive Value	Rate (95% CI)	Key Findings from Pregnancy Outcomes
Aneuploid (Positive Predictive Value)	89.2% (83.1 - 94.0)	The misdiagnosis rate after a euploid embryo transfer was 0.2% (0.0 - 0.7).
Euploid (Negative Predictive Value)	94.2% (91.1 - 96.7)	The rate for mosaic transfer, with a confirmatory euploid pregnancy outcome, was 21.7% (9.6 - 36.9).
Mosaic (PPV for confirmatory mosaic/aneuploid)	52.8% (37.9 - 67.5)	This indicates significant inaccuracy in the diagnosis of mosaicism.

The data indicates that while PGT-A is highly reliable for identifying uniform aneuploidy and euploidy, its accuracy is severely limited for mosaic embryos. The high rate of false positives among mosaics (21.7% resulted in euploid pregnancies) means that potentially viable embryos may be incorrectly deprioritized [14] [16] [12].

Key Limitations and Clinical Challenges

• Embryonic Mosaicism: Mosaicism is prevalent at the blastocyst stage and represents the most significant biological challenge to PGT-A's accuracy. A biopsy may not be representative of the entire embryo, as the trophectoderm (TE) and inner cell mass (ICM) can have different chromosomal constitutions. Furthermore, embryos can undergo "self-correction," leading to a healthy birth from an embryo diagnosed as mosaic [12].
• Technical and Sampling Errors: Errors can arise from the biopsy procedure itself, DNA amplification failure, contamination, or human error in interpretation or embryo labeling [14] [11].
• Clinical Utility Debates: The value of PGT-A as a routine screening test for all IVF patients is not conclusively demonstrated. A major multicenter randomized controlled trial (the STAR trial) found that PGT-A did not improve overall ongoing pregnancy rates per embryo transfer compared to morphology-based selection alone [15]. Professional bodies like the American Society for Reproductive Medicine (ASRM) and the American College of Obstetricians and Gynecologists (ACOG) therefore do not endorse the routine use of PGT-A for all patients, stating its best applications are still under investigation [15] [12].
• Ethical and Emotional Considerations: The procedure forces couples to make difficult decisions about discarding embryos based on probabilistic results, particularly with mosaic diagnoses. The added financial cost and the emotional impact of potentially discarding viable embryos are significant factors [11] [12].

Emerging Non-Invasive and Alternative Models

Given the limitations of PGT-A, significant research efforts are focused on developing non-invasive and artificial intelligence-based alternatives for embryo ploidy prediction.

Deep Learning Models

Deep learning (DL) models, such as the iDAScore and BELA (Blastocyst Evaluation Learning Algorithm), analyze time-lapse imaging (TLI) videos of embryo development to predict ploidy status and implantation potential without the need for biopsy [8] [17].

• iDAScore Performance: Multiple retrospective studies have shown a statistically significant but moderate association between higher iDAScore values and embryo euploidy, with Area Under the Curve (AUC) values for euploidy prediction ranging from 0.60 to 0.68 [8]. These scores are also positively associated with live birth rates and negatively associated with miscarriage rates. However, their predictive accuracy within a cohort of known euploid embryos is more modest, suggesting they are more effective as a general prioritization tool rather than a direct ploidy diagnostic [8].
• BELA Performance: The BELA model, which uses a multitask learning approach to predict a model-derived blastocyst score (MDBS) from TLI videos and then integrates maternal age, achieved an AUC of 0.76 for discriminating euploid from aneuploid embryos on its test set. In the more specific task of distinguishing euploid from complex aneuploid embryos, the AUC reached 0.83 [17]. This performance matches models trained on embryologists' manual annotations, demonstrating the potential of fully automated assessment.

Table 3: Comparison of Deep Learning Models for Ploidy Prediction

Model	Input Data	Key Performance Metric	Advantages	Limitations
iDAScore v2.0 [8]	Time-lapse video	AUC: 0.68 for euploidy	Non-invasive; can be integrated into incubator software	Moderate predictive accuracy; not a replacement for PGT-A
BELA [17]	Time-lapse video + Maternal Age	AUC: 0.76 (EUP vs. ANU)AUC: 0.83 (EUP vs. CxA)	Fully automated; requires no manual annotation; state-of-the-art performance	Performance is on a specific dataset; requires further validation

The Scientist's Toolkit: Key Research Reagents and Materials

The following table details essential materials and reagents used in PGT-A and related research, as derived from the experimental protocols cited in this review.

Table 4: Research Reagent Solutions for PGT-A and Embryo Research

Reagent / Material	Function in Protocol	Experimental Application
Laser System	To create a precise opening in the zona pellucida.	Essential for performing trophectoderm biopsy [10] [13].
Biopsy Micropipette	To aspirate and remove trophectoderm cells.	Used in conjunction with the laser for cell retrieval during TE biopsy [13].
Whole Genome Amplification (WGA) Kit	To amplify the minute quantity of genomic DNA from biopsied cells.	Mandatory pre-processing step for genetic analysis of a small cell sample [13].
Next-Generation Sequencing (NGS) Kit	For comprehensive 24-chromosome copy number analysis.	The current gold-standard platform for PGT-A analysis; also detects mosaicism [10] [15] [12].
Time-Lapse Incubator (e.g., Embryoscope+)	To culture embryos while continuously capturing images of development.	Provides the morphokinetic data required for training and deploying AI models like iDAScore and BELA [8] [17].
Febrifugine	Febrifugine|Research Compound	Febrifugine is a potent quinazolinone alkaloid with antimalarial and research applications. This product is For Research Use Only (RUO). Not for human use.
Epicorazine A	Epicorazine A, CAS:62256-05-7, MF:C18H16N2O6S2, MW:420.5 g/mol	Chemical Reagent

PGT-A, with its foundation in trophectoderm biopsy and NGS analysis, remains the gold standard for embryo ploidy assessment due to its high predictive values for uniformly euploid and aneuploid embryos. However, it is a screening tool with non-trivial limitations, most notably its invasiveness, cost, and poor accuracy in classifying mosaic embryos, which can lead to the discarding of viable embryos. The clinical evidence for its universal benefit is equivocal, and its use is not recommended for all patient populations. Emerging deep learning models like iDAScore and BELA offer promising, non-invasive alternatives for embryo prioritization. While their current diagnostic accuracy for ploidy is moderate and not yet sufficient to replace PGT-A, they represent a rapidly advancing field that may redefine the standards of embryo selection. The future of embryo ploidy prediction likely lies in integrated models that combine genetic, morphokinetic, and clinical data to maximize the safety, efficacy, and accessibility of IVF.

The comparative analysis of embryo ploidy prediction models relies heavily on understanding the evolution of cytogenetic technologies. For decades, fluorescence in situ hybridization (FISH) served as the primary method for chromosomal analysis in preimplantation genetic screening (PGS). However, its limitations in scope and resolution eventually prompted the development of more comprehensive array-based methodologies, including array comparative genomic hybridization (aCGH) and single-nucleotide polymorphism (SNP) arrays. These technological advances have fundamentally transformed reproductive medicine by enabling 24-chromosome analysis of embryos, thereby improving the accuracy of aneuploidy detection and clinical outcomes in assisted reproduction. This guide provides an objective comparison of these techniques, focusing on their performance characteristics, experimental protocols, and applications within embryo ploidy prediction research.

Technical Principles and Comparative Performance

Fluorescence In Situ Hybridization (FISH)

FISH is a cytogenetic technique that uses fluorescently labeled DNA probes to bind complementary sequences on specific chromosomes, allowing for their visualization under a fluorescence microscope [18]. The technique involves denaturing chromosomal DNA and probe DNA, followed by hybridization and signal detection [18]. In preimplantation genetic screening, FISH was traditionally applied to interphase nuclei from blastomere biopsies to assess aneuploidy for a limited number of chromosomes.

Table 1: Key Characteristics of FISH Technology

Aspect	Description
Principle	Hybridization of fluorescent DNA probes to complementary target sequences [18]
Typical Probes	Locus-specific, centromeric, or whole-chromosome painting probes [18]
Detection Method	Fluorescence microscopy [18]
Primary PGS Application	Aneuploidy screening of chromosomes 13, 15, 16, 18, 21, 22, X, and Y [19]
Key Limitation	Inability to evaluate all 24 chromosomes simultaneously [20]

Array Comparative Genomic Hybridization (aCGH)

aCGH is a microarray-based technique that detects copy number variations across the entire genome without the need for cell culture or metaphase chromosomes [21]. It works by competitively hybridizing test DNA and reference DNA, labeled with different fluorophores (e.g., Cy3 and Cy5), to thousands of DNA probes immobilized on a slide [19] [21]. The resulting fluorescence ratio at each probe location indicates relative copy number—deviations from a 1:1 ratio signify losses or gains in the test genome [21].

Single-Nucleotide Polymorphism (SNP) Arrays

SNP arrays represent a more advanced form of microarray analysis that can detect not only copy number variations but also genotype information at hundreds of thousands of single-nucleotide polymorphism sites [21] [20]. Unlike aCGH, many SNP array platforms use a single-color hybridization system where patient DNA is hybridized to the array and compared in silico to a large database of control samples [20]. This allows for the simultaneous detection of copy number changes and copy-number-neutral events like uniparental disomy (UPD) through the analysis of loss of heterozygosity (LOH) [22] [21] [20].

Table 2: Comprehensive Performance Comparison of Cytogenetic Techniques

Parameter	FISH	aCGH	SNP Array
Genome Coverage	Targeted (5-9 chromosomes typical) [23]	Comprehensive (all 24 chromosomes) [20]	Comprehensive (all 24 chromosomes) [20]
Resolution	~50 kb - 1 Mb (probe-dependent) [24]	~2.5 - 5 Mb [20]	~1.7 - 5 Mb [20]
Aneuploidy Detection	Limited to probes used [19]	All chromosomes [19]	All chromosomes [20]
Detects Segmental Aneuploidy	No (unless specifically targeted)	Yes [20]	Yes [20]
Detects UPD/LOH	No	No	Yes [22] [21] [20]
Turnaround Time	1-2 days	~12 hours [20]	~24 hours [20]
Multiplexing Capability	Limited (2-3 rounds with 5-9 probes) [23]	High (thousands of loci simultaneously)	Very High (hundreds of thousands of loci)
Throughput	Low (manual microscopy)	High	High

Experimental Data and Validation Studies

Limitations of FISH and the Transition to Comprehensive Methods

The shift from FISH to comprehensive chromosome screening (CCS) methods was driven by compelling clinical evidence. A significant limitation of FISH is its restricted chromosomal coverage, typically screening only 5-9 chromosomes despite the clinical relevance of aneuploidies in other chromosomes [19] [23]. Furthermore, studies demonstrated that FISH has a high false-positive rate; one investigation found that nearly 60% of blastocysts were chromosomally normal in multiple sections despite a cleavage-stage FISH aneuploidy diagnosis [19]. This inaccuracy stems from analyzing single cells, where technical errors or mosaicism can lead to misdiagnosis [19]. These limitations contributed to disappointing clinical outcomes in randomized controlled trials of FISH-based PGS [20].

Head-to-Head Methodological Comparisons

Direct comparative studies provide robust data on the performance of array-based platforms. In a seminal prospective double-blinded study, researchers compared aCGH and qPCR (another CCS method) by reanalyzing aCGH-diagnosed aneuploid blastocysts [19]. While 81.7% of embryos showed concordant diagnoses, 18.3% (22/120) gave discordant results for at least one chromosome [19]. Subsequent blinded reanalysis with SNP arrays revealed that the discordance was more frequently attributed to aCGH, mostly due to false positives [19]. The discordant aneuploidy call rate per chromosome was significantly higher for aCGH (5.7%) than for qPCR (0.6%) [19]. This suggests that aCGH may overdiagnose aneuploidy compared to other contemporary CCS methods.

In another comparative study focusing on hematological malignancies, SNP arrays demonstrated a significantly higher abnormality detection rate (62.5% for MDS, 72.7% for CLL) compared to aCGH (31.3% for MDS, 54.5% for CLL) and traditional cytogenetics/FISH [22]. This superior performance is largely attributed to the ability of SNP arrays to identify copy-number-neutral loss of heterozygosity (CN-LOH), which is undetectable by aCGH or FISH [22].

Figure 1: Experimental workflow and key findings from a comparative study of aCGH and qPCR for embryo ploidy assessment [19].

Detailed Experimental Protocols

Standard FISH Protocol for Embryonic Cells

The following protocol is adapted for preimplantation genetic screening on blastomere biopsies or trophectoderm samples [18]:

• Slide Preparation: Fix interphase nuclei or metaphase spreads on a microscope slide using methanol:acetic acid (3:1).
• Denaturation: Immerse slides in 70% formamide/2× SSC solution at 73°C for 5 minutes to denature chromosomal DNA.
• Dehydration: Dehydrate slides through an ethanol series (70%, 85%, 100%) for 2 minutes each and air dry.
• Probe Preparation: Prepare probe mixture according to manufacturer's instructions. Denature at 75°C for 10 minutes and pre-anneal at 37°C for 15-60 minutes.
• Hybridization: Apply probe to the target area on the slide, cover with a coverslip, seal with rubber cement, and incubate in a humidified chamber at 37°C for 6-16 hours.
• Post-Hybridization Washes: Remove coverslips and wash slides in 0.4× SSC/0.3% NP-40 at 73°C for 2 minutes, followed by 2× SSC/0.1% NP-40 at room temperature for 1 minute.
• Counterstaining and Detection: Apply DAPI counterstain and mount with antifade solution.
• Signal Analysis: Visualize using a fluorescence microscope with appropriate filter sets. Score signals in at least 10-15 interphase nuclei per probe set.

aCGH Workflow for Embryo Ploidy Assessment

The standard protocol for aCGH in comprehensive chromosome screening involves [19] [21] [20]:

• Whole Genome Amplification (WGA): Amplify genomic DNA from the embryonic biopsy (typically 5-10 trophectoderm cells) using a method such as SurePlex or similar.
• DNA Labeling: Label patient DNA with one fluorophore (e.g., Cy5) and sex-matched reference DNA with another (e.g., Cy3) using a BioPrime DNA Labeling Kit or similar.
• Hybridization: Mix labeled test and reference DNA, denature, and co-hybridize to a microarray slide (e.g., BlueGnome 24sure+ platform) for 4-16 hours.
• Washing and Scanning: Wash slides to remove non-specifically bound DNA and scan using a microarray scanner (e.g., Agilent scanner).
• Data Analysis: Analyze fluorescence ratio data using dedicated software (e.g., BlueFuse Multi). Chromosomal regions with log2 ratios significantly deviating from zero indicate copy number variations.

Figure 2: Generalized aCGH workflow for comprehensive chromosome screening of embryos, from biopsy to diagnosis.

SNP Array Protocol for PGS

The protocol for SNP array analysis shares initial steps with aCGH but diverges in labeling and analysis [20]:

• WGA: Amplify genomic DNA from the embryonic biopsy.
• Restriction Digestion: Digest amplified DNA with restriction enzymes (e.g., NspI for Affymetrix platform).
• Adapter Ligation: Ligate adapters to the digested fragments.
• PCR Amplification: Perform PCR to amplify adapter-ligated fragments.
• Fragmentation and Labeling: Fragment, label, and hybridize the PCR products to the SNP array (e.g., Affymetrix GeneChip).
• Washing and Scanning: Wash, stain, and scan the array according to manufacturer's instructions.
• Bioinformatic Analysis: Analyze data using specialized software (e.g., Genotyping Console, Nexus Copy Number) for both copy number analysis (using signal intensity) and genotype calling (using allele discrimination). This dual analysis enables detection of CNVs, UPD, and LOH.

Research Reagent Solutions and Essential Materials

Table 3: Key Research Reagents for Cytogenetic Techniques

Reagent/Material	Function	Example Applications
Fluorescently Labeled Probes (FISH)	Bind to complementary DNA sequences for visualization [18]	Locus-specific aneuploidy screening (e.g., chromosomes 13, 18, 21, X, Y)
Nick Translation DNA Labeling Kit	Enzymatically incorporates labeled nucleotides into DNA probes [23]	Generating custom FISH probes; labeling DNA for aCGH
Whole Genome Amplification Kit	Amplifies entire genome from small DNA samples [20]	aCGH and SNP analysis from single cells or small biopsies
CGH/SNP Microarray Platform	Solid support with immobilized DNA probes for genome-wide hybridization [25] [21]	Comprehensive aneuploidy screening; copy number variation detection
Cy3 and Cy5 Fluorescent Dyes	Differential labeling of test and reference DNA samples [26] [21]	aCGH experiments
Bioinformatic Analysis Software	Analyzes fluorescence ratios and genotype calls to identify abnormalities [25]	Interpreting aCGH and SNP array data; distinguishing pathological CNVs from benign variants

The evolution from FISH to array-based technologies represents a paradigm shift in embryo ploidy prediction models. While FISH provided the initial foundation for preimplantation genetic screening, its technical limitations—particularly restricted chromosomal coverage and inability to detect copy-number-neutral events—have rendered it largely obsolete for comprehensive aneuploidy screening. Array-based methodologies (aCGH and SNP arrays) offer superior genome-wide resolution, higher throughput, and demonstrated improvements in diagnostic accuracy. Among array platforms, SNP arrays provide the unique advantage of detecting uniparental disomy and loss of heterozygosity, in addition to copy number variations. The selection of an appropriate platform depends on the specific research objectives, with considerations for resolution requirements, need for genotype information, and throughput capabilities. As the field advances, these array-based technologies continue to refine our understanding of embryonic aneuploidy and improve clinical outcomes in assisted reproductive technology.

In vitro fertilization (IVF) success hinges on selecting embryos with the highest reproductive potential. For decades, preimplantation genetic testing for aneuploidy (PGT-A) using trophectoderm (TE) biopsy has been the gold standard for identifying chromosomally normal (euploid) embryos prior to transfer [27]. While effective, this approach is inherently invasive, requiring the physical removal of cells from the blastocyst, which raises concerns about potential embryo harm, technical demands, and diagnostic inaccuracies due to mosaicism [27] [28]. These limitations have catalyzed a significant drive within reproductive medicine toward developing non-invasive alternatives that can maintain diagnostic accuracy while eliminating physical intervention on the embryo.

The rationale for this shift is multifaceted. Invasive biopsy is a technically complex procedure that requires extensive training and could potentially compromise embryo viability and implantation potential [28]. Furthermore, because a TE biopsy samples only a subset of cells, it may not represent the complete genetic constitution of the embryo, leading to misdiagnosis in mosaic embryos where both euploid and aneuploid cells coexist [27] [29]. Non-invasive preimplantation genetic testing (niPGT) aims to overcome these challenges by analyzing embryonic cell-free DNA (cfDNA) passively released into the spent embryo culture medium (SCM), offering a safer and potentially more representative profile of the embryonic genome [27] [29]. Concurrently, artificial intelligence (AI) models are emerging as a completely different class of non-invasive tools that leverage time-lapse imaging and morphological data to predict ploidy status [30] [17] [31]. This guide provides a comparative analysis of these promising non-invasive technologies, evaluating their performance, methodologies, and clinical applicability against the conventional invasive standard.

Comparative Performance Analysis of Non-Invasive Technologies

The following table summarizes the performance metrics of key non-invasive ploidy prediction methods as reported in recent scientific literature.

Table 1: Performance Comparison of Non-Invasive Ploidy Prediction Technologies

Technology	Reported Concordance with TE Biopsy or AUC	Key Strengths	Major Limitations
niPGT-A (using cfDNA)	73.1% - 93.8% concordance (varies by study protocol) [29] [28]	- Truly biopsy-free [27]- Safer for the embryo [28]- Potentially profiles entire embryo [29]	- Maternal DNA contamination [27] [29]- Variable cfDNA yield & quality [27]- Challenges detecting mosaicism/segmental aneuploidies [27]
LIFE Predict v1.1 (ML Model)	AUC: 0.818 - 0.824 for predicting aneuploidy/live birth [30]	- Uses routine morphokinetic data [30]- Strong risk stratification (13.3% to 76.4% aneuploidy across score quartiles) [30]	- Does not directly assess genetics- Performance inferior to PGT-A [32]
BELA (AI Model)	AUC: 0.76 for euploid vs. aneuploid discrimination [17]	- Fully automated; no embryologist input [17]- Analyzes time-lapse sequences [17]	- Performance is dataset-dependent [31]
iDAScore v2.0 (AI Model)	AUC: 0.68 for euploidy prediction [8]	- Integrated into time-lapse incubators [8]- Also predicts live birth [8]	- Moderate predictive accuracy for ploidy [8]
FEMI (Foundation AI Model)	AUROC > 0.75 for ploidy prediction from images [31]	- Trained on ~18 million images [31]- Versatile (handles multiple embryology tasks) [31]	- A foundational model, requires further clinical validation [31]

Detailed Experimental Protocols and Methodologies

Protocol for Non-Invasive PGT-A (niPGT-A)

The niPGT-A workflow involves the collection, processing, and genetic analysis of cfDNA from the embryo's culture environment [27] [29].

• Embryo Culture and Media Collection: Embryos are cultured individually in sequential or single-step media in time-lapse incubators. On day 5 or 6, after blastocyst formation, the spent culture medium (SCM) is carefully collected. Some protocols also include the blastocoel fluid (BF) released after artificial collapse of the blastocyst [29].
• cfDNA Extraction and Whole-Genome Amplification (WGA): The collected SCM and BF samples contain trace amounts of fragmented cfDNA. Due to the low DNA quantity, a WGA step is critical to amplify the entire genome prior to analysis. Commonly used WGA kits include MALBAC, SurePlex, and Repli-G [28].
• Genetic Analysis and Sequencing: The amplified DNA is analyzed using Next-Generation Sequencing (NGS) to detect chromosomal imbalances. Bioinformatic pipelines then interpret the sequencing data to assign a ploidy status (euploid or aneuploid) for each embryo [27] [29].

Diagram: Experimental Workflow for niPGT-A

Protocol for AI-Based Ploidy Prediction (e.g., BELA)

AI models like BELA automate ploidy prediction by analyzing time-lapse imaging videos without the need for invasive biopsy or manual embryologist annotation [17].

• Data Acquisition and Preprocessing: Time-lapse images of developing embryos are acquired from incubators. The model processes video sequences from a key developmental window (e.g., 96 to 112 hours post-insemination). Images are cropped and resized to focus on the embryo.
• Multitask Learning for Blastocyst Score Prediction: The model's first component uses a Bidirectional LSTM (BiLSTM) architecture to analyze the image sequences and predict a model-derived blastocyst score (MDBS). This score mimics the morphological grading (ICM, TE, expansion) typically performed by embryologists.
• Ploidy Status Classification: The predicted MDBS, along with clinical data such as maternal age, is fed into a logistic regression classifier. This final step generates the prediction for the embryo's ploidy status (euploid vs. aneuploid) [17].

Diagram: BELA Model Architecture for Ploidy Prediction

Molecular Foundations of Non-Invasive Testing

The effectiveness of niPGT-A relies on the presence and quality of embryonic cfDNA in the culture medium. The release of this cfDNA is governed by several biological pathways, which also introduce technical challenges.

Diagram: Cellular Pathways of cfDNA Release in Embryos

• Apoptosis: A controlled process where caspase-activated DNases cleave DNA into small, predictable fragments (50-200 base pairs). This is a primary source of cfDNA, though it may over-represent DNA from genetically compromised cells [27].
• Necrosis: An unregulated form of cell death resulting in the release of larger, variably-sized DNA fragments into the medium [27].
• Active Secretion via Extracellular Vesicles (EVs): Embryos actively secrete DNA within exosomes and microvesicles. This EV-derived DNA is often more stable and less fragmented than apoptotic DNA, making it a potentially higher-quality source for genetic analysis [27].

These pathways contribute to a pool of cfDNA that is often fragmented and present in low quantities. A significant challenge is maternal DNA contamination, which can originate from residual cumulus cells or polar bodies, potentially leading to false-positive or false-negative aneuploidy calls [27] [29].

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 2: Key Reagents and Materials for Non-Invasive Ploidy Research

Item	Function/Application	Specific Examples
Time-Lapse Incubator	Provides continuous imaging of embryo development in stable culture conditions. Essential for collecting morphokinetic data for AI models and for timed SCM collection.	EmbryoScope (Vitrolife), Geri (Genea Biomedx) [33]
Whole-Genome Amplification (WGA) Kits	Amplifies trace amounts of cfDNA from SCM/BF to quantities sufficient for genetic analysis. Choice of kit impacts amplification bias and accuracy.	MALBAC, SurePlex, Repli-G, Picoplex [28]
Next-Generation Sequencing (NGS)	High-throughput sequencing technology used to detect chromosomal aneuploidies from amplified cfDNA or biopsied cells.	Various platforms (e.g., Illumina) [27] [28]
Cell Lysis Buffer	Used to lyse cells in TE biopsies or to stabilize cfDNA in collected SCM/BF samples prior to WGA.	Often included in commercial WGA kits [29]
AI/Software Platforms	Algorithms that analyze time-lapse images or videos to generate scores predictive of ploidy or implantation potential.	BELA [17], LIFE Predict [30], iDAScore [8], FEMI [31], MAIA [33]
TnPBI	TnPBI (2-n-propyl-4-p-tolylamino-1,2,3-benzotriazinium iodide) – RUO	TnPBI is a benzotriazinium salt for cardiovascular disease research. For Research Use Only. Not for diagnostic or therapeutic use.
Cefonicid	Cefonicid, CAS:61270-58-4, MF:C18H18N6O8S3, MW:542.6 g/mol	Chemical Reagent

The drive toward non-invasive methods for embryo ploidy assessment is a cornerstone of modern IVF research, motivated by the clinical necessity to enhance safety, accuracy, and accessibility. Both niPGT-A and AI-based models represent promising pathways toward a future without invasive embryo biopsy. Current data indicates that while niPGT-A can achieve high concordance with TE biopsy, it requires rigorous protocol optimization to overcome issues like maternal contamination [27] [29]. AI models, though not yet as accurate as genetic testing, offer a completely non-invasive and increasingly automated approach that leverages existing laboratory data [32] [31].

The future likely lies not in a single superior technology, but in integrated approaches. Combining the genetic precision of optimized niPGT-A with the morphological and developmental insights from AI models could provide a more comprehensive viability assessment than any single method [32]. Furthermore, foundation models like FEMI, trained on millions of images, hint at a future where AI's predictive power may closely rival genetic tests [31]. For researchers and clinicians, the ongoing challenge is to validate these technologies in large-scale prospective studies and standardize methodologies to fully realize the promise of non-invasive embryo selection.

The selection of embryos with the highest reproductive potential represents a central challenge in the field of assisted reproductive technology (ART). A key determinant of embryo viability is ploidy status, with euploid (chromosomally normal) embryos demonstrating significantly higher implantation potential and lower miscarriage rates compared to aneuploid embryos. While preimplantation genetic testing for aneuploidy (PGT-A) remains the gold standard for determining ploidy status, its invasive nature, cost, and technical demands have spurred the development of non-invasive artificial intelligence (AI) alternatives [8] [17]. These emerging technologies leverage time-lapse imaging and sophisticated algorithms to predict ploidy status, offering promising avenues for improving embryo selection.

In evaluating these novel approaches, understanding key performance metrics—specifically the Area Under the Curve (AUC), sensitivity, and specificity—is paramount for researchers and clinicians. These metrics provide standardized, quantitative measures to objectively compare the diagnostic accuracy and clinical utility of diverse prediction models [34]. AUC values offer a comprehensive measure of a model's ability to discriminate between euploid and aneuploid embryos across all possible classification thresholds. Sensitivity reflects the model's capacity to correctly identify euploid embryos, while specificity indicates its proficiency in recognizing aneuploid embryos [34]. This comparative analysis examines these critical metrics across the current landscape of embryo ploidy prediction models, providing researchers with a framework for methodological assessment and technological advancement.

Performance Metrics Comparison of Ploidy Prediction Models

The following table synthesizes performance data across major categories of ploidy prediction technologies, highlighting the progression from traditional manual assessments to advanced AI-driven approaches.

Table 1: Performance Metrics Comparison of Embryo Ploidy Prediction Models

Model Category	Specific Model	AUC	Sensitivity	Specificity	Key Input Data
Traditional AI Models	iDAScore v1.0 [8]	0.60–0.67	Not Reported	Not Reported	Time-lapse morphokinetics
	iDAScore v2.0 [8]	0.635–0.68	Not Reported	Not Reported	Time-lapse morphokinetics
Advanced Video-Based AI	BELA (with maternal age) [17]	0.76	Not Reported	Not Reported	Day 5 time-lapse video + Maternal age
	Visual-Temporal Contrastive Learning [35]	0.811	Not Reported	Not Reported	Time-lapse video sequences
3D Morphology + Machine Learning	Decision Tree Model [35]	0.978	Not Reported	Not Reported	Quantitative 3D parameters
	Extreme Gradient Boosting [35]	0.984	Not Reported	Not Reported	Quantitative 3D parameters

Experimental Protocols and Methodologies

Traditional and Video-Based Deep Learning Models

Models such as iDAScore and BELA represent a significant evolution in embryo assessment methodology. These systems typically employ convolutional neural networks (CNNs) trained on extensive datasets of time-lapse videos with known ploidy outcomes determined by PGT-A [8] [36]. The iDAScore algorithm, for instance, analyzes morphokinetic patterns and morphological features extracted automatically from time-lapse imaging, assigning embryos a score from 1.0 to 9.9 that correlates with euploidy likelihood [8]. These models function as fully automated systems, requiring no manual annotation by embryologists, thereby reducing subjectivity [17].

The BELA (Blastocyst Evaluation Learning Algorithm) framework introduces a sophisticated two-stage, multi-task learning approach. In its initial phase, the model processes Day 5 time-lapse videos (96–112 hours post-insemination) to predict a model-derived blastocyst score (MDBS) that encompasses inner cell mass (ICM), trophectoderm (TE), and expansion scores. This step utilizes a pre-trained spatial feature extractor and a BiLSTM (Bidirectional Long Short-Term Memory) architecture to analyze temporal developmental patterns. The second phase employs logistic regression, integrating the MDBS with maternal age as a continuous variable to generate the final ploidy prediction [17]. This methodological innovation allows BELA to leverage both morphological and clinical features, contributing to its enhanced performance with an AUC of 0.76 [17].

Table 2: Key Research Reagent Solutions for Ploidy Prediction Research

Research Tool	Primary Function	Application Context
Time-Lapse Incubators (e.g., EmbryoScope+)	Maintain stable culture conditions while capturing sequential embryo images	Provides essential morphokinetic data for deep learning model training [8] [36]
PGT-A (Preimplantation Genetic Testing for Aneuploidy)	Genetic analysis of trophectoderm biopsy samples	Establishes ground truth ploidy status for model training and validation [8] [17]
Convolutional Neural Networks (CNNs)	Automated feature extraction from embryo images/videos	Backbone architecture for most deep learning-based ploidy prediction models [36]
U-Net Architecture	Semantic segmentation of biological images	Used in 3D morphology studies for precise segmentation of TE cells and ICM [35]
SHapley Additive exPlanations (SHAP)	Interpreting machine learning model output	Identifies critical developmental timepoints influencing model predictions [17]

3D Morphology and Quantitative Machine Learning

A distinct methodological approach moves beyond conventional 2D imaging to employ 3D morphology measurement for ploidy prediction. This technique involves capturing multi-view images of Day 6 blastocysts by manually rotating them during the trophectoderm biopsy preparation phase. Using spherical rotation SIFT algorithms, these 2D images are reconstructed into a 3D model, from which quantitative morphological parameters are extracted [35].

Key parameters include trophectoderm cell number, TE cell size variance, and inner cell mass area—all of which demonstrate statistically significant differences between euploid and non-euploid blastocysts. These quantitative parameters serve as inputs for various machine learning models, including decision trees and extreme gradient boosting (XGBoost) classifiers [35]. This approach achieves remarkable performance, with AUC values reaching 0.984, while offering superior model interpretability compared to deep learning "black box" systems. The quantitative criteria extracted from these models provide biologically plausible insights, indicating that euploid blastocysts typically exhibit higher trophectoderm cell counts, larger ICM area, and reduced TE cell size variance [35].

Visualizing Experimental Workflows

The following diagrams illustrate the core methodologies and logical relationships underlying the primary ploidy prediction approaches discussed in this analysis.

Workflow for Video-Based Deep Learning Models

Video-Based AI Prediction Pipeline - This workflow depicts the standard process for video-based deep learning models like BELA, showing the integration of image data and clinical features.

Workflow for 3D Morphology-Based Prediction

3D Morphology Prediction Pipeline - This diagram outlines the 3D morphology-based approach, highlighting its strength in generating interpretable quantitative criteria.

This comparative analysis reveals a clear performance hierarchy among ploidy prediction methodologies. Traditional AI models like iDAScore demonstrate moderate predictive capability (AUC 0.60-0.68), serving as useful adjuncts for embryo prioritization but lacking the accuracy required to replace PGT-A [8]. Advanced video-based approaches like BELA show improved discrimination (AUC 0.76-0.81) by leveraging comprehensive temporal data and integrating clinical variables like maternal age [17] [35]. Most impressively, 3D morphology with machine learning achieves exceptional performance (AUC >0.97) through precise quantification of structural parameters, while offering superior interpretability [35].

These metrics underscore a fundamental trade-off between model complexity, interpretability, and performance. While 3D approaches currently deliver superior accuracy, their requirement for specialized imaging presents implementation challenges. Video-based systems offer a practical balance of performance and feasibility for clinical integration. For researchers, the selection of appropriate performance metrics—AUC for overall discriminative capacity, plus sensitivity and specificity for clinical utility at specific thresholds—remains essential for rigorous model validation. Future advancements will likely focus on multi-modal approaches that combine the strengths of these methodologies, ultimately enhancing objective embryo assessment and improving IVF outcomes.

Methodological Innovations: AI Algorithms and Morphokinetic Analysis

The selection of embryos with the highest reproductive potential remains a central challenge in in vitro fertilization (IVF). Preimplantation genetic testing for aneuploidy (PGT-A) is the gold standard for identifying chromosomally normal (euploid) embryos but is invasive, costly, and not universally applicable [8]. Deep learning models offer a promising, non-invasive alternative by analyzing time-lapse imaging (TLI) to predict embryo ploidy status and viability. This guide provides a comparative analysis of three prominent deep learning models—BELA, iDAScore, and STORK-A—focusing on their architectures, training methodologies, and performance in embryo ploidy prediction, to inform researchers and drug development professionals in the field of reproductive medicine.

Model Architectures and Training Approaches

BELA: Blastocyst Evaluation Learning Algorithm

BELA employs a multi-step, fully automated pipeline that uniquely combines model-predicted blastocyst scores with maternal age for ploidy prediction [17].

• Architecture: BELA uses a multitask learning approach. The first component processes day-5 time-lapse videos (96–112 hours post-insemination, hpi) using a pre-trained spatial feature extractor and a BiLSTM (Bidirectional Long Short-Term Memory) model to concurrently predict inner cell mass (ICM), trophectoderm (TE), expansion, and a overall blastocyst score, generating a Model-Derived Blastocyst Score (MDBS) [17]. The second component uses this MDBS along with maternal age in a logistic regression model to predict ploidy status [17].
• Training Data: The model was trained and evaluated on internal datasets from Weill Cornell Medicine (WCM), comprising 1998 Embryoscope and 841 Embryoscope+ time-lapse sequences, with PGT-A results as ground truth [17].
• Key Innovation: BELA eliminates the need for manual embryologist annotations by automatically predicting blastocyst scores directly from time-lapse sequences [17].

iDAScore: Intelligent Data Analysis Score

iDAScore is a deep learning-based scoring system designed for fully automated embryo evaluation and ranking based on the likelihood of clinical pregnancy or fetal heartbeat [37] [38].

• Architecture: The model is a convolutional neural network (CNN). Its input consists of 128 frames, sampled at one-hour intervals on a single focal plane (resolution 256x256 pixels), covering embryo development from 12 to 140 hours post-insemination [38]. The network outputs a score from 1.0 to 9.9.
• Training Data: iDAScore is trained on an exceptionally large and diverse dataset. Its algorithm has been developed using full time-lapse sequences of over 180,000 embryos with known clinical outcomes, incorporating data from multiple continents to minimize bias related to patient profiles and clinical protocols [37].
• Key Innovation: iDAScore provides a rapid, objective ranking of all a patient's embryos without requiring manual annotations, significantly improving workflow efficiency in the IVF laboratory [37] [38].

STORK-A: Ploidy Prediction from Static Images

STORK-A is a machine learning algorithm developed to predict embryo ploidy status from a single static image captured at 110 hours post-insemination [17].

• Architecture: While specific architectural details of STORK-A are not elaborated in the provided results, it is positioned as a predecessor to more advanced models like BELA. It represents an earlier approach that relies on image-based analysis rather than full time-lapse sequences [17].
• Training Data: The model was trained using datasets from Weill Cornell Medicine [17].
• Performance Context: BELA, a more advanced model, has been shown to surpass STORK-A's performance in ploidy prediction tasks [17].

Performance Comparison and Experimental Data

The table below summarizes the key performance metrics of the featured models in ploidy prediction, based on available validation studies.

Table 1: Performance Comparison of Deep Learning Models in Ploidy Prediction

Model	Primary Function	Key Performance Metrics (Ploidy Prediction)	Training Dataset Size	Validation Notes
BELA [17]	Ploidy & Quality Prediction	AUC: 0.76 (EUP vs. ANU, with maternal age) [17]	1,998 + 841 sequences [17]	Multitask learning; outperforms STORK-A [17]
iDAScore v2.0 [8]	Embryo Viability Scoring	AUC: 0.68 (for euploidy prediction) [8]	>180,000 time-lapse sequences [37]	Correlates with live birth; large-scale validation [8] [37]
STORK-A [17]	Ploidy Prediction	Surpassed by BELA model [17]	WCM datasets [17]	Predecessor model using single images [17]
FEMI [9]	Foundation Model (Multiple Tasks)	AUROC >0.75 (Image-based ploidy prediction) [9]	~18 million time-lapse images [9]	A more recent, large-scale foundational model for comparison.

FEMI is included as a state-of-the-art reference; it is a foundational model trained on ~18 million images that achieves high performance across multiple embryology tasks, including ploidy prediction [9].

Table 2: Summary of Clinical Correlation and Key Advantages

Model	Correlation with Clinical Outcomes	Key Advantages
BELA [17]	Predicts blastocyst score and uses it for ploidy classification.	Fully automated; no embryologist input required; integrates maternal age.
iDAScore [37] [38]	Significantly correlated with live birth rates (p<0.001) [38]. OR for Live Birth: 1.81 (95% CI: 1.67-1.98) [37].	High throughput, objective ranking; saves embryologist time; large, diverse training set.
STORK-A [17]	Provides a baseline for image-based ploidy prediction.	Simpler architecture relying on single time-point images.

Experimental Protocols and Methodologies

BELA's Model Training and Validation Workflow

The development of BELA followed a structured, multi-dataset approach to ensure robustness and generalizability [17].

• Data Preparation: Time-lapse sequences from WCM's Embryoscope and Embryoscope+ incubators were used. Each sequence typically contained 360-420 distinct frames captured at 0.3-hour intervals over 5 days of development. PGT-A results served as the ground truth for ploidy status [17].
• Model Training: The model was trained using a four-fold cross-validation on the WCM-Embryoscope dataset. An ablation analysis was conducted to identify the most informative developmental time points (96-112 hpi) for prediction [17].
• Validation: BELA's performance was evaluated on internal test sets (WCM-Embryoscope and WCM-Embryoscope+) as well as external datasets from IVI Valencia (Spain) and IVF Florida to assess its generalizability [17]. Performance was gauged using accuracy, AUC, precision, and recall.

iDAScore Validation and Clinical Workflow

External validation studies for iDAScore, such as the one conducted at Tongji Hospital, demonstrate its real-world clinical application and correlation with live birth outcomes [38].

• Study Design: A large retrospective cohort study analyzed 6,291 single vitrified-thawed blastocyst transfer cycles. Blastocysts were cultured in an EmbryoScope+ incubator and retrospectively scored by the iDAScore model [38].
• Data Analysis: Blastocysts were sorted by their iDAScore values and divided into four groups for comparison. Outcomes measured included clinical pregnancy, miscarriage, and live birth rates. Uni- and multivariable logistic regressions were performed to assess the correlation between iDAScore and live birth [38].
• Key Findings: The study confirmed that iDAScore was significantly correlated with clinical pregnancy, miscarriage, and live birth (p < 0.001), providing external validation of its effectiveness in a clinical setting [38].

The Scientist's Toolkit: Essential Research Reagents and Materials

The development and application of these deep learning models rely on a foundation of specific laboratory protocols, reagents, and hardware. The table below details key components of the experimental ecosystem.

Table 3: Essential Research Reagents and Materials for Model Development

Item / Solution	Function / Role	Example Use Case
Time-Lapse Incubator	Provides undisturbed embryo culture and continuous image acquisition for generating training data.	EmbryoScope+ system used for culturing embryos and capturing time-lapse sequences [38] [36].
Culture Media	Supports embryo development in vitro.	G-TL (Vitrolife) or SAGE (Origio) media used in embryo culture protocols [39].
PGT-A Kits & Reagents	Provides ground truth data for ploidy status for model training and validation.	VeriSeq PGS kit (Illumina) used for ploidy analysis of biopsied samples [39].
Image Segmentation Tools	Pre-processes raw embryo images to isolate the embryo from the background, improving model input quality.	U-NET architecture used for blastocyst image segmentation before CNN-based model development [39].
GPU-Accelerated Computing	Enables efficient training of complex deep learning models on large image datasets.	Training of iDAScore was performed using Nvidia Quadro RTX8000 GPUs [38].
4-Hydroxybaumycinol A1	Rubeomycin B|Anthracycline Antibiotic|For Research	Rubeomycin B is an anthracycline antibiotic for cancer research. It inhibits DNA replication. For Research Use Only. Not for human use.
Tiformin	Tiformin, CAS:4210-97-3, MF:C5H12N4O, MW:144.18 g/mol	Chemical Reagent

The comparative analysis of BELA, iDAScore, and STORK-A reveals distinct architectural philosophies and trade-offs. BELA demonstrates the power of a fully automated, multi-task pipeline that integrates model-derived quality scores with clinical features like maternal age. iDAScore stands out for its massive, diverse training dataset and proven clinical utility in predicting live birth, offering significant gains in laboratory efficiency. STORK-A represents an important foundational approach using static images. While not a replacement for PGT-A, these models show moderate to strong predictive power and offer a non-invasive, scalable, and objective method for embryo assessment. Future advancements will likely involve even larger foundation models like FEMI and prospective randomized trials to further solidify their role in clinical practice [9] [30].

In vitro fertilization (IVF) success hinges on selecting the single most viable embryo for transfer, a complex challenge in reproductive medicine. Traditional embryo selection primarily relies on static morphological assessment at isolated time points, an approach limited by subjectivity, inherent inter-observer variability, and the disruption of stable culture conditions [40] [41]. The emergence of time-lapse imaging (TLI) systems has introduced a paradigm shift, enabling continuous, non-invasive monitoring of embryonic development within stable incubator environments. This technology provides an uninterrupted sequence of images, capturing the dynamic morphokinetic parameters of development—the precise timing of key embryonic events [40] [42]. The subsequent critical step is feature extraction: the process of quantifying these developmental sequences into actionable data for predicting embryo viability and ploidy status. This guide provides a comparative analysis of the methodologies and technologies bridging TLI data and clinical decision-making, with a specific focus on their application in predicting embryo ploidy to improve IVF outcomes.

Comparative Analysis of Ploidy Prediction Models

Extracted features from TLI sequences are used to train various predictive models. The table below compares the performance, methodology, and key characteristics of leading ploidy prediction models as identified in recent literature.

Table 1: Comparative Analysis of Embryo Ploidy Prediction Models

Model Name	Model Type	Key Input Features	Reported AUC for Ploidy Prediction	Strengths	Limitations/Challenges
BELA [17]	Deep Learning (Multitask)	Entire time-lapse video (96-112 hpi); Maternal age	0.76 (EUP vs. ANU, with age)	Fully automated; no manual annotation; uses full video context.	Performance is dataset-dependent; requires significant computational resources.
iDAScore (v1.0 & v2.0) [8]	Deep Learning (CNN-based)	Time-lapse videos with known outcomes	0.60 - 0.68 (for euploidy)	Integrated into clinical workflows (EmbryoScope+); scores correlate with live birth.	Modest predictive accuracy for ploidy; not a replacement for PGT-A.
LIFE Predict v1.1 [30]	Machine Learning (Ensemble)	Morphokinetic meta-variables (Range, MAEkinetic); clinical data	0.818 (External Validation)	Quantifies deviation from optimal development; strong risk stratification.	Requires precise morphokinetic annotation; prospective validation needed.
STORK-A [17]	Machine Learning	Single static image (110 hpi)	~0.74 (from cited literature)	Simplicity of using a single time point.	Lacks dynamic developmental context.
ERICA [17]	Deep Learning	Single static embryo images	0.74	Demonstrated early feasibility of AI for ploidy prediction.	Lower sensitivity (54%); limited by static image input.

Experimental Protocols for Feature Extraction and Model Validation

Data Acquisition and Pre-processing

The foundational step for any analysis is the generation of high-quality, standardized TLI data. Embryos are cultured in integrated time-lapse incubators (e.g., EmbryoScope+ or Eeva system) that capture high-resolution images at frequent intervals (e.g., every 5-20 minutes) over 5-7 days without removing them from stable culture conditions [41]. The resulting datasets are substantial, often comprising 360-420 distinct frames per embryo [17]. Key pre-processing steps include:

• Data Annotation: Embryologists may annotate the videos with precise timings for specific morphokinetic events, such as the time of division to 2-cells (t2), 3-cells (t3), and so on, based on established consensus [40].
• Ground Truth Establishment: For ploidy prediction models, the ground truth is typically established via Preimplantation Genetic Testing for Aneuploidy (PGT-A), which involves trophectoderm biopsy and chromosomal analysis [8] [17].
• Data Curation: Studies often use retrospective datasets, splitting them into training/testing and external validation sets to assess model generalizability. For instance, the LIFE Predict v1.1 model was trained on 833 embryos and externally validated on 357 embryos from different clinics [30].

Protocol for Traditional Morphokinetic Parameter Analysis

This protocol involves the manual or semi-automated extraction of specific time intervals from the TLI sequences.

• Objective: To identify correlations between the timing of early cleavage events and embryo ploidy or implantation potential.
• Methodology:
  • Define t0: The start time of development is set, typically as the time of insemination for IVF or sperm injection for ICSI [40].
  • Annotation: Record the exact time (post-t0) when the embryo reaches each cell stage (t2, t3, t4, t5, t8, etc.).
  • Calculate Durations: Derive parameters such as the duration of the 2-cell stage (cc2, t3-t2), the synchrony of the second and third cell cycles (s2, t4-t3), and the time to complete the third cleavage (cc3, t8-t4 or t5-t3) [40].
  • Statistical Analysis: Compare these parameters between outcome groups (e.g., implanted vs. non-implanted, euploid vs. aneuploid) to identify predictive thresholds.
• Example Findings: Studies have shown that embryos that implant often cleave faster, reaching the 2-cell, 3-cell, and 4-cell stages sooner than those that do not. Parameters like t4 and s2 have been reported to differ between euploid and aneuploid embryos [40].

Protocol for Deep Learning-Based Feature Extraction

This modern approach uses neural networks to automatically extract relevant features directly from the image data, without relying on pre-defined morphokinetic parameters.

• Objective: To leverage the full information content of TLI videos for end-to-end prediction of ploidy or viability.
• Methodology (as exemplified by BELA [17]):
  • Input: Processed time-lapse video from a key developmental window (e.g., 96-112 hours post-insemination).
  • Spatial Feature Extraction: A pre-trained convolutional neural network (CNN) processes individual frames to generate feature vectors representing the image content at each time point.
  • Temporal Integration: A Bidirectional Long Short-Term Memory (BiLSTM) network analyzes the sequence of feature vectors to understand the temporal dynamics of development.
  • Multitask Learning: The model is trained to simultaneously predict blastocyst quality scores (Inner Cell Mass, Trophectoderm, Expansion) and the final ploidy status, which improves feature robustness.
  • Ploidy Prediction: The model-derived blastocyst score is combined with maternal age in a final logistic regression layer to output a ploidy prediction.

The following workflow diagram illustrates the typical process for a deep learning-based analysis of time-lapse imaging data.

The Scientist's Toolkit: Essential Reagents and Materials

Successful implementation of TLI analysis requires a suite of specialized laboratory equipment, software, and reagents. The following table details the key components of this research and clinical toolkit.

Table 2: Key Research Reagent Solutions for TLI Analysis

Item Name	Type	Primary Function in TLI Analysis
EmbryoScope+/EmbryoScope [41]	Time-Lapse Incubator System	Provides integrated, stable culture conditions while capturing high-resolution images at set intervals without disturbing embryos.
Eeva System [41]	Time-Lapse Incubator System	Automatically analyzes early-stage morphokinetic parameters (first 48 hours) to generate a viability score.
PGT-A Kits & Reagents [17]	Genetic Test Consumables	Provide the ground truth for embryo ploidy status against which TLI-based prediction models are trained and validated.
Specialized Culture Media	Laboratory Reagent	Supports embryo development over the extended 5-7 day culture period within the TLI system.
iDAScore Software [8]	AI Analysis Algorithm	A deep learning model integrated into EmbryoScope+ that analyzes time-lapse videos to assign an embryo score (1.0-9.9) correlating with implantation potential and euploidy.
Generative AI Models [43]	Data Augmentation Tool	Generates synthetic embryo images to address data scarcity, augment training datasets, and improve the robustness of deep learning classifiers.
Epibetulinic acid	Epibetulinic Acid|TGR5 Agonist for Research	Epibetulinic acid is a potent TGR5 agonist for metabolic disease and inflammation research. This product is For Research Use Only. Not for human or veterinary diagnostic or therapeutic use.
Dimethyl peroxide	Dimethyl peroxide, CAS:690-02-8, MF:C2H6O2, MW:62.07 g/mol	Chemical Reagent

The comparative analysis reveals a spectrum of methodologies for extracting features from developmental sequences, each with distinct advantages. Traditional morphokinetic parameter analysis provides a transparent, clinically intuitive framework but may miss subtler patterns captured by deep learning models like BELA and iDAScore. These AI-driven approaches demonstrate promising but moderate accuracy (AUCs largely between 0.60-0.82) in predicting ploidy, confirming they are not yet a replacement for PGT-A [8] [17] [41]. The future of this field lies in several key areas: the standardization of protocols and algorithms across clinics to improve generalizability, the prospective validation of models like LIFE Predict v1.1 in real-world settings, and the integration of TLI features with other non-invasive biomarkers such as secreted factors or metabolic profiles [41] [30]. Furthermore, techniques to overcome data scarcity, including the use of synthetic data generation [43] and federated learning, will be crucial for developing more robust and generalizable models. For researchers and clinicians, the choice of feature extraction method must be guided by the specific clinical question, available resources, and a clear understanding that these technologies serve best as powerful, non-invasive adjuncts for embryo prioritization rather than definitive diagnostic tools.

The pursuit of reliable, non-invasive methods to identify viable embryos represents a central challenge in assisted reproductive technology (ART). Traditional embryo selection has largely relied on static morphological assessment, where embryologists grade embryos based on visual characteristics at specific developmental stages. However, the subjective nature and limited predictive power of these methods have driven the development of more quantitative approaches. The advent of time-lapse imaging (TLI) systems enabled the detailed tracking of embryonic development, generating rich morphokinetic data—the precise timings of key developmental events [36].

Morphokinetic meta-variables represent a sophisticated evolution beyond simple event timing. They are computational constructs that quantify patterns and deviations across the entire developmental trajectory of an embryo. Rather than analyzing individual milestones in isolation, these meta-variables synthesize complex temporal information to provide a holistic assessment of developmental dynamics [30]. This analytical framework is increasingly important for predicting embryo ploidy status (chromosomal normalcy), a critical determinant of implantation success and live birth outcomes. This guide provides a comparative analysis of how morphokinetic meta-variables perform against other embryo assessment methodologies within the rapidly advancing field of embryo ploidy prediction research.

Comparative Analysis of Ploidy Prediction Methodologies

Fundamental Approaches to Embryo Assessment

Table 1: Core Methodologies in Embryo Ploidy Prediction

Methodology	Primary Data Input	Key Principle	Automation Level	Key Advantage
Traditional Morphology	Static blastocyst images	Visual grading of morphology (ICM, TE, expansion)	Manual	Widespread availability, low technical barrier
Basic Morphokinetics	Timings of specific events (t2, t3, tSB, etc.)	Correlation between delayed development and aneuploidy	Semi-automated	Adds dynamic temporal dimension to assessment
Video-Based Deep Learning (e.g., BELA)	Raw time-lapse video sequences	End-to-end feature learning from pixel data	Fully automated	Eliminates subjectivity of manual annotation
Morphokinetic Meta-Variables (e.g., LIFE Predict)	Calculated trajectory deviations (Range, MAEkinetic)	Quantification of developmental path deviation from an optimal model	Fully automated	Holistic pattern recognition of entire developmental journey

Performance Comparison of Predictive Models

Recent research has generated quantitative performance data for various ploidy prediction approaches, allowing for direct comparison of their discriminatory power.

Table 2: Quantitative Performance Metrics of Ploidy Prediction Models

Model / Approach	Reported AUC (95% CI)	Dataset Size	Key Predictors	Study
LIFE Predict v1.1 (Meta-variables)	0.824 (0.806-0.868)	1,190 embryos	Morphokinetic meta-variables, clinical data	Güell et al., 2025 [30]
BELA (Video-Based DL)	0.76 (maternal age included)	1,998 sequences	Time-lapse videos (96-112 hpi), maternal age	Nature Communications, 2024 [17]
iDAScore v2.0 (Commercial DL)	0.68 (p<0.001)	249,635 embryos	Time-lapse video features	Bori et al., 2025 [8]
Logistic Regression (Mixed Effects)	0.71 (0.67-0.73)	8,147 embryos	Morphokinetic timings, blastocyst grade	Bamford et al., 2023 [44]
Morphokinetics Only Model	0.61	8,147 embryos	Morphokinetic timings alone	Bamford et al., 2023 [44]
Embryo Grading Only Model	0.52	8,147 embryos	Traditional morphology grades alone	Bamford et al., 2023 [44]

The data reveal a clear performance hierarchy. Models incorporating morphokinetic meta-variables and advanced deep learning consistently outperform traditional statistical models using basic morphokinetic parameters. Most notably, traditional morphological grading alone shows minimal discriminatory power for ploidy status (AUC ≈ 0.52), underscoring the critical limitation of conventional assessment methods [44].

Experimental Protocols and Methodologies

The LIFE Predict v1.1 Model: A Meta-Variable Approach

The LIFE Predict v1.1 model exemplifies the application of morphokinetic meta-variables. Its development followed a rigorous experimental protocol:

Dataset Composition: A retrospective multicentre cohort study utilized 1,190 blastocysts from nine fertility clinics, with confirmed outcomes (either live birth or PGT-A diagnosis). The dataset was split with 70% (n=833) for model training and testing, and 30% (n=357) for external validation [30].

Core Meta-Variable Calculation: The model's innovation lies in two novel meta-variables:

• Range: Quantifies the overall deviation of an embryo's morphokinetic timings from expected values observed in embryos that resulted in live births.
• MAEkinetic (Mean Absolute Error kinetic): Measures the average absolute deviation across all morphokinetic data points, providing a comprehensive measure of developmental alignment with optimal trajectories [30].

Model Training and Validation: An ensemble machine learning model was trained using these meta-variables combined with clinical data. Performance was assessed via cross-validation and external validation using AUC-ROC metrics. The model's clinical utility was further evaluated by stratifying aneuploidy risk across score quartiles and within standard morphological grades [30].

The BELA Model: An End-to-End Deep Learning Approach

The Blastocyst Evaluation Learning Algorithm (BELA) represents an alternative methodology that bypasses manual feature engineering:

Architecture Design: BELA employs a two-stage, multitask learning framework. The first component processes day-5 time-lapse videos (96-112 hours post-insemination) using a pre-trained spatial feature extractor and a BiLSTM network to predict blastocyst score components (ICM, TE, expansion) directly from pixel data [17].

Input Processing: The model takes complete time-lapse sequences as input, transformed into feature vectors. Unlike meta-variable approaches, BELA autonomously identifies critical developmental time points, with SHAP analysis revealing heightened importance at approximately 96 hpi and 112 hpi [17].

Ploidy Prediction: In the second stage, the model-derived blastocyst score (MDBS) is combined with maternal age in a logistic regression classifier to predict ploidy status. This approach achieved an AUC of 0.76 for discriminating between euploid and aneuploid embryos when maternal age was included [17].

Visualizing Methodological Relationships

The diagram below illustrates the conceptual relationships and workflow differences between the major approaches to embryo ploidy prediction.

Diagram 1: Ploidy Prediction Methodologies compares how different approaches process time-lapse data, showing the progression from manual assessment to automated meta-variables and deep learning.

The performance hierarchy evident in Table 2 can be visualized through the following relationship diagram.

Diagram 2: Ploidy Prediction Performance Hierarchy illustrates how predictive accuracy improves with methodological sophistication, from basic morphology to advanced meta-variables.

The Scientist's Toolkit: Essential Research Reagents and Platforms

Table 3: Key Research Tools for Morphokinetic Embryo Assessment

Tool / Technology	Primary Function	Research Application	Example Implementation
Time-Lapse Incubators (EmbryoScope+)	Continuous embryo imaging in stable culture conditions	Generates raw morphokinetic data for analysis	Platform for iDAScore integration [8]
Preimplantation Genetic Testing for Aneuploidy (PGT-A)	Chromosomal status determination via trophectoderm biopsy	Provides ground truth for model training and validation	Used in all cited studies for outcome verification [30] [17]
Computer Vision Models (EfficientNet-V2, ResNet-3D)	Automated image feature extraction from time-lapse data	Enables frame-by-frame developmental stage classification	Achieved 87% accuracy for 17 morphokinetic stages [45]
Sequence Learning Architectures (BiLSTM)	Temporal pattern recognition in sequential data	Analyzes developmental trajectories across timepoints	Core component of BELA model for blastocyst score prediction [17]
Meta-Variable Algorithms (Range, MAEkinetic)	Quantification of developmental trajectory deviations	Calculates holistic measures of embryo developmental normalcy	LIFE Predict v1.1's novel contribution [30]
SHAP (SHapley Additive exPlanations)	Model interpretability and feature importance	Identifies critical developmental timepoints	Used in BELA to reveal importance peaks at 96hpi and 112hpi [17]
Tasuldine	Tasuldine|C10H9N3S|CAS 88579-39-9	Tasuldine is a bronchosecretolytic research agent. This product is for research use only (RUO) and is not intended for personal use.	Bench Chemicals
Dethiophalloidin	Dethiophalloidin|Phalloidin Analog|For Research		Bench Chemicals

The comparative analysis demonstrates that morphokinetic meta-variables represent a significant methodological advancement in non-invasive embryo ploidy prediction. By quantifying developmental trajectories holistically rather than focusing on isolated timings, these constructs achieve superior predictive performance (AUC 0.824) compared to both traditional methods and other computational approaches [30].

For research applications, meta-variables offer the distinct advantage of providing quantifiable, standardized metrics of developmental normality that can be correlated with molecular mechanisms of chromosomal segregation errors. The consistent inverse relationship between LIFE Predict scores and aneuploidy rates across quartiles (76.4% to 13.3%) provides a robust experimental framework for investigating the phenotypic expression of aneuploidy [30].

Future research directions should focus on prospective validation of these technologies in diverse clinical settings, integration with multi-omics data to establish biological correlates of abnormal developmental trajectories, and development of more sophisticated meta-variables that capture non-linear developmental patterns. As these tools evolve, they promise to not only improve clinical embryo selection but also to serve as valuable research platforms for understanding the fundamental biology of early human development and the mechanisms underlying embryonic aneuploidy.

The selection of embryos with the highest potential for achieving a successful pregnancy is a paramount objective in assisted reproductive technology (ART). Two of the most critical and widely available parameters for embryo selection are maternal age and embryo morphology. While each factor provides valuable standalone information, a growing body of evidence demonstrates that their integrated analysis offers a more powerful, synergistic approach for predicting embryonic viability and ploidy status. This comparative analysis examines the individual and combined predictive value of these clinical parameters, situating them within the broader context of emerging non-invasive ploidy prediction technologies, particularly artificial intelligence (AI)-based models. Understanding the interplay between traditional morphological assessment, maternal age, and next-generation predictive algorithms is essential for researchers and clinicians aiming to optimize embryo selection protocols and improve in vitro fertilization (IVF) outcomes.

Comparative Analysis of Predictive Values

The Independent and Combined Predictive Power of Morphology and Maternal Age

Clinical studies consistently demonstrate that both embryo morphology and maternal age independently influence pregnancy outcomes, even when chromosomally normal (euploid) embryos are transferred.

Table 1: Impact of Euploid Blastocyst Morphology on Pregnancy Outcomes

Embryo Morphology Grade	Sustained Implantation Rate (Age 33)	Sustained Implantation Rate (Age 39)	Adjusted Odds Ratio (aOR) for Live Birth
Day 5 Good Quality	86%	80%	Reference (1.00)
Day 5 Fair Quality	71%	62%	Not Specified
Day 5 Poor Quality	59%	55%	Not Specified
Day 6 Blastocysts (All Qualities)	81%	46%	Not Specified
Inner Cell Mass (ICM) Grade C	Not Specified	Not Specified	0.32 (p=0.03)

Data synthesized from [46] [47]

A study analyzing 610 natural-cycle frozen euploid embryo transfers (NC-FET) found that blastocyst morphology significantly impacts pregnancy and live birth rates. Specifically, euploid blastocysts with an inner cell mass (ICM) graded as "C" had statistically significant decreased odds of achieving a clinical pregnancy and live birth compared to those with an ICM grade "A" [46]. Another retrospective analysis of 229 transferred euploid embryos confirmed that good quality day 5 euploid blastocysts had the highest sustained implantation rates (80-90%) across all maternal ages, outperforming fair, poor, and day 6 blastocysts [47].

Table 2: Independent Impact of Maternal Age on Outcomes with Top-Quality Euploid Embryos

Maternal Age Group	Clinical Pregnancy and Live Birth Rates with AA-Graded Euploid Blastocysts
< 35 years	Highest Rates
35-39 years	Intermediate Rates
40+ years	Lowest Rates

Data synthesized from [46]

Critically, maternal age remains an independent predictor of success even when a top-graded (AA) euploid embryo is transferred [46]. This suggests that age-related factors, potentially of endometrial origin, continue to influence implantation and gestation, even after the chromosomal barrier has been overcome.

Performance Comparison with AI-Based Ploidy Prediction Models

Artificial intelligence models represent a paradigm shift in non-invasive embryo assessment, often utilizing the very parameters of morphology and maternal age but analyzing them in novel, data-driven ways.

Table 3: Performance Comparison of AI Models in Predicting Embryo Ploidy

AI Model / Approach	Input Data	Key Performance Metric	Performance Value
BELA Model [17]	Time-lapse videos (96-112 hpi) + Maternal Age	AUC (EUP vs. ANU)	0.76
BELA Model [17]	Time-lapse videos (96-112 hpi) + Maternal Age	AUC (EUP vs. CxA)	0.826
End-to-End Deep Learning [48]	Raw time-lapse videos (Days 1-5)	AUC (ANU vs. EUP/Mosaic)	0.74
Gradient Boosting (HOG+PCA) [49]	Processed static blastocyst images	Aneuploid Recall	0.84
Meta-Analysis (Pooled) [50]	Various embryonic imaging	Pooled Sensitivity / Specificity	0.71 / 0.75

Abbreviations: AUC, Area Under the Receiver Operating Characteristic Curve; EUP, Euploid; ANU, Aneuploid; CxA, Complex Aneuploid; hpi, hours post-insemination; HOG, Histogram of Oriented Gradients; PCA, Principal Component Analysis.

The BELA (Blastocyst Evaluation Learning Algorithm) model exemplifies the modern integration of clinical parameters. It is a multi-task learning model that first predicts a blastocyst score from time-lapse videos and then uses this model-derived blastocyst score (MDBS) in conjunction with maternal age to predict ploidy status [17]. This approach achieved an AUC of 0.76 in discriminating between euploid and aneuploid embryos, matching the performance of models trained on embryologists' manual scores [17]. A comprehensive meta-analysis of 20 studies confirmed the promising performance of AI, with a summary AUC of 0.80 for predicting embryonic euploidy based on imaging data [50].

Experimental Protocols and Methodologies

Protocol for Clinical Outcome Studies (Morphology & Age)

The evidence for the integration of morphology and maternal age is largely derived from rigorous retrospective cohort studies.

• Patient Selection and Embryo Transfer: Studies typically include patients undergoing single euploid blastocyst transfers in a natural or hormone replacement cycle. Key exclusion criteria often involve uterine abnormalities, advanced endometriosis, and the use of donor oocytes to isolate the effect of maternal age [46].
• Blastocyst Grading: Embryos are graded according to the Gardner and Schoolcraft system prior to vitrification. This system assesses blastocyst expansion stage and the morphology of the inner cell mass (ICM) and trophectoderm (TE), with grades from A (best) to D (worst) [46].
• Ploidy Assessment: Trophectoderm biopsy is performed on day 5 or 6 blastocysts. Biopsied cells are analyzed using comprehensive chromosomal screening techniques, such as next-generation sequencing (NGS), to determine euploidy or aneuploidy [46] [49].
• Statistical Analysis: Multivariable logistic regression models are employed to adjust for potential confounders such as maternal age at transfer, nulliparity, and endometrial lining thickness. Generalized estimating equations (GEE) may be used to account for correlations between patients with multiple transfer cycles [46].

Protocol for AI-Based Ploidy Prediction Models

AI model development follows a structured pipeline for image processing, feature extraction, and model training.

• Data Curation and Preprocessing: Time-lapse videos or static images of embryos with known PGT-A results are collected. The dataset is split into training and testing sets (e.g., 80:20). Image preprocessing is critical and may include:
  • Image Segmentation: Isolating the embryo from the background and removing interfering objects (e.g., pipettes) [49].
  • Augmentation: Techniques like rotation, color enhancement, and blurring are applied to increase dataset diversity and robustness [49].
  • Temporal Alignment: For video-based models, frames are aligned to key developmental time points (e.g., 96-112 hours post-insemination) [17].
• Feature Extraction: This step converts images into quantifiable features.
  • Deep Learning-Based: Convolutional Neural Networks (CNNs) like VGG19 or ResNet can automatically extract spatial features from images [49].
  • Handcrafted Features: Algorithms like the Histogram of Oriented Gradients (HOG) can be used to create feature descriptors based on image gradients and orientations, followed by dimensionality reduction using Principal Component Analysis (PCA) [49].
• Model Architecture and Training:
  • Video-Based Models (e.g., BELA): Use a pre-trained CNN for spatial feature extraction from individual frames, followed by a temporal model like a Bidirectional LSTM (BiLSTM) to analyze the sequence of features over time [17].
  • Integration of Clinical Data: The model-derived morphological scores are combined with clinical variables like maternal age using a classifier such as logistic regression for the final ploidy prediction [17].
  • Validation: Model performance is rigorously evaluated on a held-out test set or via cross-validation, reporting metrics like AUC, accuracy, sensitivity, and specificity [17] [50].

Visualizing Workflows and Relationships

AI Model Workflow for Integrated Ploidy Prediction

The following diagram illustrates the end-to-end workflow of a sophisticated AI model like BELA, which integrates time-lapse imaging and clinical parameters for ploidy prediction.

Relationship Between Key Parameters and Outcomes

This conceptual diagram maps the complex interactions between maternal age, embryo morphology, ploidy status, and clinical outcomes, highlighting the role of AI integration.

The Scientist's Toolkit: Research Reagent Solutions

Table 4: Essential Materials and Reagents for Embryo Ploidy and Morphology Research

Item	Function in Research	Example Application in Context
Time-Lapse Incubator System	Provides continuous, uninterrupted culture and imaging of embryos, generating morphokinetic data.	Essential for capturing the video sequences used by AI models like BELA and for annotating precise morphokinetic parameters [17] [51].
Preimplantation Genetic Testing for Aneuploidy (PGT-A)	Gold standard for determining embryo chromosomal constitution; provides ground truth for model training.	Used to validate the ploidy status of embryos in both clinical outcome studies and as labels for supervised AI model training [17] [50] [49].
Specialized Embryo Culture Media	Supports embryo development from cleavage stage to blastocyst in vitro.	A constant in all protocols; studies use specific commercial media (e.g., SAGE Biopharma) for consistent blastocyst culture and trophectoderm biopsy [48].
Image Processing & Feature Extraction Algorithms	Convert raw embryo images into quantifiable features for analysis.	Algorithms like Histogram of Oriented Gradients (HOG) or pre-trained CNNs (VGG19, ResNet) are used to extract features for machine learning models [49].
Deep Learning Frameworks	Provide the computational architecture for building and training predictive AI models.	Used to implement complex models like CNNs for image analysis and LSTMs for temporal sequence processing of time-lapse data [17] [48].
(1S,2S)-bitertanol	(1S,2S)-Bitertanol|Chiral Fungicide
Arphamenine B	Arphamenine B, MF:C16H24N4O4, MW:336.39 g/mol	Chemical Reagent

The integration of maternal age and embryo morphology remains a cornerstone of effective embryo selection. Evidence robustly confirms that both parameters are independent yet complementary predictors of implantation and live birth success, even in the context of euploid embryo transfer. The emergence of AI-based ploidy prediction models does not render these traditional parameters obsolete; rather, it recontextualizes them. Sophisticated algorithms like BELA quantitatively automate morphological assessment and seamlessly integrate it with maternal age, achieving performance that begins to approach the predictive value of invasive PGT-A. For researchers and clinicians, the future of embryo selection lies not in choosing between traditional parameters and novel AI, but in harnessing their synergistic potential. This integrated approach promises to enhance the accuracy of non-invasive embryo viability assessment, ultimately streamlining the path to a successful pregnancy for patients undergoing ART.

The selection of viable embryos is a critical determinant of success in in vitro fertilization (IVF). A key aspect of this process is assessing embryo ploidy status—identifying chromosomally normal (euploid) embryos, which have a high likelihood of leading to a successful pregnancy, and distinguishing them from chromosomally abnormal (aneuploid) embryos, which are associated with miscarriage and failed implantation [17]. Preimplantation genetic testing for aneuploidy (PGT-A) is the current gold standard for this assessment but is invasive, costly, and not universally accessible [17] [8]. This has driven the development of non-invasive artificial intelligence (AI) models that can predict ploidy status using time-lapse imaging and clinical data.

The evolution of these models presents a compelling case study in the comparative performance of classical machine learning algorithms, such as Logistic Regression (LR), and more complex Advanced Neural Networks (ANNs). This guide provides an objective, data-driven comparison of these algorithmic approaches within the specific context of embryo ploidy prediction, summarizing experimental data and detailing methodologies to inform researchers and scientists in the field.

The following tables synthesize quantitative performance metrics from recent studies, allowing for a direct comparison of model efficacy.

Table 1: Overall Performance Metrics for Ploidy Prediction

Algorithm / Model	Task (Prediction)	AUC	Sensitivity	Specificity	Key Input Data
BELA (ANN: BiLSTM) [17]	Euploid (EUP) vs. Aneuploid (ANU)	0.76	N/A	N/A	Time-lapse videos, Maternal age
BELA (ANN: BiLSTM) [17]	Euploid (EUP) vs. Complex Aneuploid (CxA)	0.826	N/A	N/A	Time-lapse videos, Maternal age
LIFE Predict v1.1 (Ensemble ML) [30]	Aneuploidy / Live Birth	0.818	N/A	N/A	Morphokinetic meta-variables, Clinical data
iDAScore v2.0 (Deep Learning) [8]	Euploidy	0.68	N/A	N/A	Time-lapse videos
PGT-Plus (Random Forest) [52]	Abnormal Ploidy (e.g., Triploidy)	0.99 - 1.00	N/A	N/A	Ultra-low-coverage sequencing data

Table 2: Comparative Performance of Logistic Regression vs. Neural Networks

Study Context	Logistic Regression Performance	Advanced Neural Network Performance
Feature: Model-Derived Blastocyst Score (MDBS)Task: EUP vs. ANU prediction [17]	AUC: ~0.66 (using MDBS)AUC: ~0.76 (using MDBS + maternal age)	BELA (BiLSTM) generated the MDBS from time-lapse videos, which was then used in the LR model for ploidy classification.
Feature: Morphokinetic Meta-VariablesTask: Aneuploidy prediction [30]	Performance was compared against an ensemble model (Random Forest, XGBoost). LR was part of the model comparison during development.	The final ensemble model (LIFE Predict v1.1), which may have incorporated LR as a component, achieved an AUC of 0.824.
Feature: Genomic DataTask: Ploidy abnormality identification [52]	One of three models tested.	Random Forest achieved superior performance (AUC ~1.0) compared to SVM and Logistic Regression.

Experimental Protocols & Methodologies

The BELA Model: A Hybrid Workflow

The Blastocyst Evaluation Learning Algorithm (BELA) exemplifies a sophisticated hybrid methodology that leverages both neural networks and logistic regression [17].

Workflow: BELA Model for Ploidy Prediction

• Input Data Preparation: The model uses time-lapse sequences of embryo development, typically comprising 360-420 frames captured over 5 days. For ploidy prediction, BELA focuses on a specific window of 96-112 hours post-insemination (hpi) [17].
• Spatial Feature Extraction: Processed video frames are fed into a pre-trained spatial feature extraction model (typically a Convolutional Neural Network, or CNN) to convert raw images into informative feature vectors [17].
• Temporal Modeling with BiLSTM: These feature vectors are processed by a multitasking Bidirectional Long Short-Term Memory (BiLSTM) network. This advanced neural network architecture is designed to model temporal sequences and concurrently predict sub-scores for inner cell mass (ICM), trophectoderm (TE), expansion, and the overall blastocyst score (BS) [17].
• Ploidy Classification with Logistic Regression: The model-derived blastocyst score (MDBS) is then used as the primary input feature for a Logistic Regression classifier. Notably, this LR model also incorporates maternal age as a continuous input feature to finalize the ploidy status prediction (EUP vs. ANU) [17].

The LIFE Predict v1.1 Model: An Ensemble Approach

This model employs a distinct strategy centered on novel morphokinetic meta-variables [30].

• Input Data and Preprocessing: The model was trained on a multicenter dataset of 1,190 blastocysts. The primary data sources were time-lapse imaging annotations and accompanying clinical data.
• Feature Engineering - Meta-variables: The core innovation involves two novel meta-variables designed to quantify deviations from normative development patterns observed in embryos that resulted in live births:
  • Range: Measures the spread of an embryo's morphokinetic parameters relative to a live birth reference.
  • MAE_kinetic: Calculates the Mean Absolute Error between an embryo's morphokinetic parameters and the expected values from the live birth model [30].
• Model Training and Architecture: The LIFE Predict v1.1 is an ensemble model. During development, the performance of multiple machine learning algorithms, including Logistic Regression, Support Vector Machine (SVM), and Random Forest, was compared. The final model integrates the strengths of these algorithms to achieve robust performance [30].

The PGT-Plus Model: A Genomic Data Approach

This model addresses ploidy prediction from a different angle, using genomic data from preimplantation genetic testing [52].

• Input Data: The model utilizes ultra-low-coverage whole-genome sequencing (ulc-WGS) data from embryo biopsies.
• Feature Extraction: From the sequencing data, 23 continuous candidate features are derived, including heterozygosity rates and likelihood ratios of alleles across chromosomes.
• Model Selection and Training: After feature selection using Gini importance analysis, three classifiers were trained and compared: Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression. The study concluded that the Random Forest model demonstrated superior performance for this specific task and was selected as the final PGT-Plus model [52].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials and Tools for Embryo Ploidy Prediction Research

Item	Function in Research	Example Use Case
Time-Lapse Incubator	Provides a stable culture environment while capturing continuous images of embryo development at set intervals.	EmbryoScope/EmbryoScope+ used to generate the time-lapse sequences for models like BELA and iDAScore [17] [8].
Preimplantation Genetic Testing for Aneuploidy (PGT-A)	Serves as the gold standard for establishing ground-truth labels of embryo ploidy status for model training and validation.	Used in all cited studies to confirm euploidy or aneuploidy in the embryos used in the datasets [17] [30].
Spatial Feature Extractor (e.g., CNN)	A pre-trained deep learning model that processes raw embryo images to identify and extract salient morphological features.	The first step in the BELA pipeline, converting images into feature vectors for the BiLSTM [17].
Recurrent Neural Network (e.g., BiLSTM)	A type of neural network architecture specialized for sequential data; capable of learning from the entire time-lapse sequence.	Used in BELA to analyze the temporal sequence of embryo features and predict quality scores [17].
Morphokinetic Meta-Variables	Computed metrics that quantify an embryo's developmental trajectory against a normative model.	The core features (Range, MAEkinetic) in the LIFE Predict v1.1 model that encapsulate deviation from optimal development [30].
1-Hydroxy-2-butanone	1-Hydroxybutan-2-one (CAS 5077-67-8)|Endogenous Metabolite

The comparative analysis within the niche field of embryo ploidy prediction reveals a nuanced landscape. Advanced Neural Networks, particularly architectures like BiLSTMs and CNNs, excel at automatically learning complex, non-linear patterns from high-dimensional data such as raw time-lapse videos. Their strength lies in feature extraction and modeling temporal dynamics without heavy reliance on manual annotation [17].

Conversely, Logistic Regression remains a powerful, interpretable tool for classification tasks when provided with high-quality, engineered features. Its performance is strongly dependent on the input features it receives, as demonstrated by its role in the BELA model where it effectively combined the neural network-derived blastocyst score with maternal age [17].

The prevailing trend leans toward hybrid and ensemble approaches. These methodologies leverage the strengths of both paradigms: using ANNs for automated feature discovery from complex data, and employing either LR or other classic models (like Random Forest) for robust final classification based on those features and other clinical variables [17] [30]. This synergy, rather than a head-to-head competition, appears to be the most promising path forward for developing robust, clinically applicable AI tools in reproductive medicine.

Clinical Implementation Challenges and Model Optimization Strategies

Embryonic mosaicism, the presence of two or more chromosomally distinct cell lines within a single embryo, presents a significant challenge in assisted reproductive technology (ART). The accurate detection and interpretation of mosaicism are critical for embryo selection, yet current methodologies exhibit substantial limitations that impact clinical decision-making. This comparative analysis examines the performance of leading mosaic detection platforms, evaluating their technical capabilities, diagnostic accuracy, and clinical applicability within the framework of embryo ploidy prediction research.

The prevalence of mosaicism in human embryos is remarkably high, with single-cell sequencing revealing that 100% of blastocysts exhibit some degree of chromosomal mosaicism [53]. This finding fundamentally challenges traditional embryo selection paradigms and underscores the need for refined detection methodologies. As ART laboratories increasingly implement preimplantation genetic testing for aneuploidy (PGT-A), understanding the technical limitations of various detection platforms becomes essential for both clinical application and research advancement.

Methodological Approaches and Technical Principles

Conventional PGT-A Methodologies

Traditional PGT-A approaches utilize trophectoderm (TE) biopsy followed by next-generation sequencing (NGS) to assess chromosomal status. The standard laboratory workflow involves blastocyst culture, TE biopsy at day 5-6 of development, whole genome amplification, and NGS-based copy number variation analysis [54] [27]. Embryos are typically classified as euploid, aneuploid, or mosaic based on established thresholds, with mosaicism commonly defined when copy number values fall within the 20-80% range between monosomy and disomy or between disomy and trisomy [54].

Table 1: Standard PGT-A Classification Thresholds

Classification	Copy Number Threshold	Typical Clinical Interpretation
Euploid	<20% abnormal cells	Recommended for transfer
Mosaic	20-80% abnormal cells	Case-by-case evaluation
Aneuploid	>80% abnormal cells	Not recommended for transfer

A significant advancement in conventional PGT-A is the implementation of dual classification systems. Recent studies propose categorizing mosaicism into Mosaic-A (conventional mosaic embryos in standard reports) and Mosaic-B (includes both Mosaic-A and aneuploid embryos containing mosaic features), providing a more comprehensive framework for understanding mosaicism biological behavior [54].

Single-Cell Sequencing Approaches

Single-cell sequencing methodologies represent the most precise approach for mosaicism detection, enabling karyotype analysis at individual cell resolution. The experimental protocol involves complete embryo digestion, mechanical separation of all visible cells, whole-genome sequencing at approximately 0.3× depth per cell, and copy number variation analysis across all chromosomes [53]. This approach allows for direct distinction between meiotic and mitotic error origins through analysis of aneuploidy distribution patterns across all embryonic cells.

The key advantage of single-cell methodologies is their ability to detect "chromosome-complementary" cells (where one cell shows chromosome gain while another shows loss of the same chromosome), observed in approximately 70% of blastocysts [53]. This phenomenon, undetectable by bulk analysis methods, demonstrates how conventional multicell biopsies significantly underestimate true mosaicism prevalence.

Non-Invasive PGT-A (niPGT) Platforms

Non-invasive approaches analyze cell-free DNA (cfDNA) released into spent embryo culture medium, eliminating biopsy requirements. The niPGT protocol involves embryo culture for 5-6 days, collection of spent medium, cfDNA extraction, whole genome amplification, and NGS analysis [27]. The molecular basis of cfDNA release involves multiple pathways including apoptosis (producing 50-200bp fragments via caspase-activated DNases), necrosis, active DNA secretion through extracellular vesicles, and chromatin remodeling processes [27].

Despite its non-invasive advantage, niPGT faces significant technical challenges including variable cfDNA yield, potential maternal DNA contamination, and sequencing biases that impact detection accuracy, particularly for mosaic and segmental aneuploidies [27].

Artificial Intelligence-Based Assessment

Deep learning algorithms offer a completely non-invasive alternative by analyzing time-lapse imaging data. Platforms such as BELA (Blastocyst Evaluation Learning Algorithm) utilize convolutional neural networks to process time-lapse videos, employing multitask learning to predict blastocyst scores which are then integrated with maternal age for ploidy prediction [17]. The iDAScore system represents another AI approach, applying deep learning to time-lapse videos to assign scores from 1.0 to 9.9 based on developmental patterns correlated with ploidy status [8].

These systems typically analyze specific developmental windows (96-112 hours post-insemination) identified as most predictive through ablation studies, with feature importance analysis revealing bimodal distribution patterns aligned with embryological assessment criteria [17].

Comparative Performance Analysis

Detection Accuracy and Diagnostic Concordance

Comprehensive benchmarking of mosaic variant calling strategies reveals significant methodological variability in detection capabilities. A systematic evaluation of 11 mosaic detection approaches based on a whole-exome reference standard containing 354,258 control positive mosaic single-nucleotide variants demonstrated condition-dependent performance variations across platforms [55].

Table 2: Mosaic Detection Algorithm Performance Metrics

Algorithm Category	Representative Tools	SNV Detection AUC	INDEL Detection Performance	Optimal VAF Range
Mosaic-specific	MosaicForecast, DeepMosaic	0.60-0.68	Moderate (F1 score: 0.55-0.65)	5-35%
Modified somatic	Mutect2 (tumor-only)	0.65-0.72	Low (F1 score: 0.45-0.55)	4-25%
Modified germline	HaplotypeCaller (ploidy-adjusted)	0.58-0.64	Moderate-high at VAF ≥16%	16-50%
Ensemble approaches	M2S2MH	0.68-0.75	Variable	Full spectrum

For mosaic single-nucleotide variants (SNVs), MosaicForecast and Mutect2 tumor-only mode demonstrated superior performance in low to medium variant allele frequency (VAF) ranges (4-25%), while mosaic-specific algorithms outperformed in higher VAF ranges (>25%) [55]. The evaluation noted substantial discordance between different algorithms, with variant call agreement rarely exceeding 32% between different methodological approaches.

Mosaicism Detection in Clinical PGT-A

Clinical PGT-A data from large-scale analyses reveals significant variability in mosaicism reporting across platforms and laboratories. A study of 36,506 blastocysts found an overall mosaicism rate of 23% using standard classification, with significant maternal age-dependent patterns [54]. The proportion of mosaic embryos classified as Mosaic-A decreased with advancing maternal age (31% in women <35 years to 10% in women >42 years), while the broader Mosaic-B classification demonstrated an opposite trend, increasing from 46% to 62% across the same age groups [54].

Another analysis of 86,208 embryos from 17,366 patients reported an overall mosaicism rate of 15.8%, with stratification revealing higher rates of low-level mosaicism (20-40%) and segmental abnormalities in younger patients, while older patients exhibited increased high-level mosaicism (40-80%) and complex whole-chromosome abnormalities [56]. These findings highlight how detection methodology influences observed age-related patterns in mosaicism prevalence.

Limitations in Mosaicism Detection

All current methodologies face significant limitations in accurate mosaicism detection:

TE Biopsy Limitations: Conventional TE biopsy suffers from sampling error, typically assessing only 5-10 cells from the trophectoderm, potentially missing abnormal cell lines present in other embryonic regions. The diagnostic concordance between TE biopsy and single-cell analysis is substantially limited, with one study revealing discordance rates exceeding 70% for specific chromosomal abnormalities [53].

niPGT Technical Challenges: Non-invasive approaches demonstrate moderate-to-high concordance with TE biopsy (typically 70-85%), but exhibit reduced sensitivity for detecting mosaicism and segmental aneuploidies due to technical limitations including DNA degradation artifacts, variable cfDNA representation, and inability to distinguish embryonic from maternal DNA contamination [27].

AI-Based Prediction Limitations: Deep learning models show moderate predictive value for ploidy status, with area under the curve (AUC) values ranging from 0.60-0.76 for euploidy prediction [8] [17]. However, these systems cannot differentiate specific aneuploidy types or mosaic patterns, and their performance remains insufficient for standalone diagnostic application without genetic testing confirmation.

Experimental Protocols and Workflows

Comprehensive Single-Cell Sequencing Protocol

Sample Preparation: Blastocysts are completely digested using protease-based enzymatic treatment, followed by mechanical dissociation using mouth pipetting to generate single-cell suspensions. All visible cells are individually collected under microscopic visualization [53].

Whole Genome Sequencing: Individual cells undergo low-coverage (0.3×) whole genome sequencing using multiple displacement amplification for whole genome amplification. Library preparation utilizes tagmentation-based approaches for efficient DNA fragment generation [53].

Copy Number Variation Analysis: Sequencing data is processed using computational pipelines that normalize read counts across genomic bins, detect significant deviations from expected diploid ratios, and assign confidence scores for aneuploidy calls. The variability score thresholding excludes cells with aberrantly high scores (>5.38% of cells) potentially affected by amplification biases [53].

Data Interpretation: Meiotic-origin aneuploidies are defined when ≥95% of cells display uniform aneuploidy. Mitotic aneuploidies are identified through heterogeneous distribution patterns across the embryonic cell population. Phylogenetic reconstruction utilizes complementary chromosome patterns to infer developmental timing of mitotic errors [53].

Benchmarking Framework for Mosaic Variant Calling

Reference Standard Design: The benchmarking platform employs 39 mixtures of six pre-genotyped normal cell lines, creating mosaic simulations with known variant allele frequencies (0.5-56%). This generates 354,258 control positive mosaic SNVs/INDELs and 33,111,725 control negatives across three mixture categories (M1, M2, M3) representing distinct lineage relationships [55].

Performance Evaluation: Algorithms are evaluated across multiple conditions including VAF spectrum (0.5-56%), sequencing depth (125× to 1,100×), variant types (SNVs/INDELs), and variant sharing patterns. Performance metrics include precision-recall curves, F1 scores, and false positive rates per megabase [55].

Condition-Specific Optimization: The benchmarking identifies optimal algorithm selection based on specific research requirements: MosaicForecast excels for low-VAF SNVs (<10%), HaplotypeCaller with ploidy adjustment performs best for medium-to-high VAF ranges (>25%), and ensemble approaches provide the most comprehensive detection across diverse VAF spectra [55].

Research Reagent Solutions

Table 3: Essential Research Reagents for Mosaicism Detection Studies

Reagent/Platform	Manufacturer	Primary Application	Technical Specifications
VeriSeq PGS Kit	Illumina	NGS-based PGT-A	24-chromosome screening, ≥20Mb resolution
SurePlex DNA Amplification System	Illumina	Whole genome amplification	Efficient amplification from single cells
EmbryoScope+ Time-Lapse System	Vitrolife	AI-based embryo assessment	Continuous imaging without culture disturbance
iDAScore Software	Vitrolife	Deep learning embryo scoring	Algorithm trained on >249,635 embryo videos
MiSeq System	Illumina	NGS sequencing	Mid-output sequencing for PGT-A applications
BlueFuse Multi Software	Illumina	PGT-A data analysis	Automated aneuploidy calling and mosaicism assessment

Visualization of Methodological Workflows

Diagram 1: Comparative Workflows for Mosaicism Detection Methodologies. Three primary approaches demonstrate varying technical complexities and capability profiles, with distinct limitations and advantages for research applications.

The comprehensive analysis of embryonic mosaicism detection methodologies reveals a complex landscape of complementary technologies, each with distinctive capabilities and limitations. Traditional TE biopsy with NGS provides clinical utility but suffers from inherent sampling constraints and resolution limitations. Single-cell sequencing approaches offer unprecedented resolution for research applications but remain impractical for routine clinical use. Emerging technologies including niPGT and AI-based assessment show promising non-invasive potential but require further validation and refinement.

Future methodological development should focus on integrated approaches that combine the precision of single-cell analysis with the clinical applicability of non-invasive platforms. The establishment of standardized benchmarking frameworks, such as the mosaic variant calling reference standard, will enable systematic improvement of detection algorithms across diverse methodological platforms. As evidence increasingly demonstrates the developmental potential of mosaic embryos, refined detection capabilities will play a crucial role in optimizing embryo selection and advancing reproductive outcomes.

The integration of artificial intelligence (AI) into in vitro fertilization (IVF) represents a paradigm shift in embryo selection, moving beyond traditional morphological assessment. The accurate prediction of embryo ploidy (chromosomal normality) is a critical determinant of IVF success, as euploid embryos have a significantly higher potential for successful implantation and live birth [8] [9]. However, the development of robust and clinically reliable ploidy prediction models faces two fundamental challenges: multi-center variability in data and inconsistencies in embryo annotation. This guide provides a comparative analysis of contemporary AI models, focusing on their performance across diverse datasets and the methodologies employed to ensure annotation consistency.

Comparative Performance of Embryo Ploidy Prediction Models

The performance of AI models can vary significantly based on their architecture, training data, and specific tasks. The table below summarizes the key performance metrics of several prominent models as reported in multi-center studies.

Table 1: Performance Comparison of AI Models for Embryo Assessment

Model Name	Primary Task	Reported Performance (Metric)	Data Variability & Key Finding
iDAScore (v1 & v2) [8]	Euploidy prediction	AUC: 0.60 - 0.68 (across 6 studies)	Performance is consistent but moderate across multiple centers; more effective when ploidy status is unknown.
FEMI [9] [57]	Euploidy prediction	AUROC > 0.75	Significantly outperforms benchmark models; trained on ~18 million images from multiple clinics.
MAIA [33]	Clinical pregnancy prediction	Overall Accuracy: 66.5%; AUC: 0.65	Developed for a specific population (Brazil), highlighting impact of demographic diversity on model performance.
Single Instance Learning (SIL) CNNs [58]	Live-birth prediction / Rank ordering	AUC ~0.60; Kendall's W: ~0.35	Exhibits high rank-order instability and critical error rates (~15%) across different fertility centers.
Automated Morphokinetic Model [59]	Morphokinetic stage detection	Accuracy: 87% (17 stages)	Aims to standardize the annotation of developmental timings, reducing a key source of inter-observer variability.

Experimental Protocols and Methodologies

A critical understanding of model performance requires insight into their training and evaluation protocols.

Foundational Model Training (FEMI)

FEMI (Foundational IVF Model for Imaging) utilizes a self-supervised learning (SSL) approach, which is a key differentiator from models trained solely on labeled data [9] [57].

• Architecture: Vision Transformer Masked Autoencoder (ViT MAE).
• Training Data: Approximately 18 million time-lapse images from multiple clinics (e.g., Weill Cornell Medicine, IVI RMA Valencia, IVF Florida) and public datasets.
• Pre-training: The model was first pre-trained on ImageNet-1k, then on the large, unlabeled collection of time-lapse images. The SSL task was to reconstruct original images from masked inputs, allowing the model to learn domain-specific features without manual annotation.
• Downstream Task Fine-tuning: For ploidy prediction, the FEMI encoder was fine-tuned using sequences of time-lapse images (video input from 96-112 hours post-insemination) and incorporated maternal age as a predictive feature. This was evaluated against benchmark models like VGG16, EfficientNet V2, and MoViNet.

Stability Assessment of Conventional Models

A landmark study systematically evaluated the stability of Single Instance Learning (SIL) Convolutional Neural Networks (CNNs), which are commonly used in research and commercial platforms [58].

• Experimental Design: Fifty replicate models with identical architecture and training data were generated, varying only in the random initialization "seed."
• Datasets: Models were trained on a dataset from Massachusetts General Hospital (MGH) and tested on an external dataset from Weill Cornell.
• Evaluation Metrics:
  • Rank Consistency: Measured using Kendall’s coefficient of concordance (Kendall's W) to assess agreement in embryo rankings across the 50 replicate models.
  • Critical Error Rate: The frequency at which a low-quality (e.g., degenerate) embryo was ranked highest when a viable blastocyst was available.
  • Interpretability Analysis: Used Gradient-weighted Class Activation Mapping (Grad-CAM) and t-SNE to visualize the divergent decision-making strategies of different replicates.

Standardizing Morphokinetic Annotation

Inconsistent manual annotation of morphokinetic events is a major source of data variability. To address this, one study developed a highly accurate machine learning model for automating this process [59].

• Model Architecture: EfficientNet-V2-Large, fine-tuned on a publicly available dataset of 273,438 labeled embryo images.
• Input and Post-processing: The model used static greyscale images. A novel post-processing algorithm was applied to reduce prediction noise and precisely detect the exact timing of each morphokinetic stage change.
• Validation: The model was tested on an independent dataset, achieving an F1-score of 0.881 and an accuracy of 87% across 17 morphokinetic stages, representing a significant improvement over previous models.

The following workflow diagram illustrates the contrasting approaches between traditional models and modern foundational models like FEMI in handling data variability.

The Scientist's Toolkit: Essential Research Reagents and Materials

The development and validation of embryo ploidy prediction models rely on a suite of specialized tools and technologies.

Table 2: Key Research Reagent Solutions for AI-Based Embryo Assessment

Item / Technology	Function in Research & Development
Time-Lapse Incubators (e.g., EmbryoScope+) [8] [60]	Provides the primary source data: continuous, non-invasive imaging of embryo development without disturbing culture conditions.
Preimplantation Genetic Testing for Aneuploidy (PGT-A) [8] [9]	Serves as the "ground truth" for model training and validation by providing the definitive ploidy status of each embryo.
Vision Transformer (ViT) Models [9]	A modern neural network architecture effective at capturing complex patterns in large-scale image datasets, used by foundational models like FEMI.
Convolutional Neural Networks (CNNs) [58] [60]	The traditional and widely used deep learning architecture for image analysis tasks, forming the basis of many commercial and research models.
Gradient-weighted Class Activation Mapping (Grad-CAM) [58]	An interpretability tool that produces visual explanations for decisions made by CNN-based models, helping to identify features influencing predictions.
Self-Supervised Learning (SSL) Frameworks [9]	Allows models to pre-train on vast amounts of unlabeled data to learn general features of embryo development before fine-tuning on specific, labeled tasks.

The comparative analysis reveals a clear trajectory in the evolution of embryo ploidy prediction models. While established tools like iDAScore provide consistent, moderate performance, they and other conventional CNNs are hampered by significant multi-center variability and instability in clinical tasks like rank-ordering [8] [58]. The emergence of foundational models like FEMI, trained on massive, diverse datasets using self-supervised learning, points toward a more robust and standardized future [9]. Furthermore, the automation of morphokinetic annotation is a crucial step in resolving the persistent challenge of annotation consistency [59]. For researchers and clinicians, this underscores that the choice of model must be informed not only by headline performance metrics but also by rigorous, multi-center validation of its stability and generalizability.

The selection of embryos with the highest reproductive potential is a cornerstone of successful in vitro fertilization (IVF). Traditionally, preimplantation genetic testing for aneuploidy (PGT-A) has been the gold standard for assessing embryonic ploidy status, a critical factor for implantation and live birth. However, PGT-A is invasive, costly, and raises ethical considerations [8] [17]. In recent years, artificial intelligence (AI) algorithms have emerged as promising non-invasive alternatives for embryo selection, leveraging time-lapse imaging and deep learning to predict ploidy and viability [8] [31] [17].

A paramount challenge in the clinical deployment of these AI models is generalizability—their ability to maintain robust performance across diverse fertility clinics and heterogeneous patient populations. Variability in laboratory protocols, culture conditions, and patient demographics (e.g., maternal age distributions) can significantly impact model performance [61] [58]. This comparative analysis evaluates the generalizability of leading embryo ploidy prediction models, synthesizing experimental data on their cross-clinic performance, methodological approaches to mitigating bias, and overall reliability.

Performance Comparison of Embryo Ploidy Prediction Models

Quantitative performance metrics across different models and validation settings are summarized in the table below. The Area Under the Receiver Operating Characteristic Curve (AUC) is a key metric, where a value of 1.0 indicates perfect prediction and 0.5 indicates performance no better than chance.

Table 1: Performance Metrics of Embryo Ploidy Prediction Models

Model Name	Reported AUC (Primary Validation)	Externally Validated AUC	Key Input Data	Maternal Age Included?	Primary Outcome
FEMI [31]	0.76	Data not available	~18 million time-lapse images	Yes	Ploidy Status
BELA [17]	0.76 (WCM-Embryoscope)	0.66 (Spain dataset)	Time-lapse sequences (96-112 hpi)	Yes	Ploidy Status
iDAScore v1.0 [8]	0.60 - 0.68 (for euploidy)	Data not available	Time-lapse sequences	No	Fetal Heartbeat / Euploidy
AI Models (Pooled) [62]	0.80 (Pooled)	Data not available	Embryonic images (various)	Various	Ploidy Status
STORK-A [17]	0.74	Data not available	Single image (110 hpi)	No	Ploidy Status
ERICA [17]	0.74	Data not available	Embryo images	No	Ploidy Status

The data reveals a performance range for ploidy prediction, with top models like FEMI and BELA achieving AUCs of approximately 0.76 on their internal tests [31] [17]. A meta-analysis of 20 studies found that AI models have a pooled AUC of 0.80 for predicting embryonic euploidy, with a sensitivity of 0.71 and specificity of 0.75 [62]. However, performance can be more modest and variable in external validation cohorts; for instance, BELA's AUC decreased from 0.76 to 0.66 when tested on an external dataset from Spain [17]. Furthermore, a study on iDAScore v1.0 highlighted that clinic-specific AUCs for predicting fetal heartbeat varied substantially from 0.58 to 0.69 before accounting for different maternal age distributions between clinics [61].

Experimental Protocols for Assessing Generalizability

Robust evaluation of model generalizability relies on specific experimental designs and statistical methods. Key methodologies cited in the literature include:

External Validation Across Multiple Clinics

The most direct method for assessing generalizability is external validation, where a model developed on data from one or more "source" clinics is tested on a completely separate dataset from one or more "target" clinics. For example, the BELA model was trained on data from Weill Cornell Medicine (WCM) and then tested on independent datasets from IVI Valencia and IVF Florida [17]. This process reveals how well a model's learned features translate to new clinical environments with different equipment, protocols, and patient populations. Performance metrics are compared between internal and external tests to quantify the performance drop.

Age-Standardization of Performance Metrics

Maternal age is a powerful confounder in embryo viability. To isolate a model's performance from the effects of varying age distributions between clinics, researchers have developed a method for age-standardizing the Area Under the Curve (AUC) [61]. This involves:

• Defining a common reference age population.
• For each clinic, calculating weights for each embryo based on the relative frequency of its maternal age in the clinic's population compared to the reference population.
• Calculating a weighted ROC curve (WROC) and its associated AUC (AUC~WROC~) using these weights.

This method was shown to reduce between-clinic variance in AUC by 16%, enabling a more direct comparison of the model's intrinsic discriminatory power across sites [61].

Evaluation of Model Stability and Rank-Order Consistency

Beyond predictive accuracy, the stability of model outputs is critical for clinical reliability. One study evaluated this by training 50 replicate convolutional neural networks with identical architectures and training data but different random initializations ("seeds") [58]. They then assessed the consistency of embryo rank-ordering for individual patients across all models using Kendall's coefficient of concordance (W). The study found poor consistency (Kendall's W ≈ 0.35) and high critical error rates (≈15%), where low-quality embryos were incorrectly ranked as the top choice [58]. This indicates that some AI models may produce unstable and inconsistent recommendations, undermining their clinical reliability.

The Generalizability Challenge: A Conceptual Workflow

The following diagram illustrates the workflow from model development to the key challenges and methods for assessing generalizability across diverse clinical settings.

Research Reagent Solutions for Embryo Ploidy AI Research

The development and validation of generalizable AI models require specific data, software, and hardware components. The table below details key resources as identified in the surveyed literature.

Table 2: Essential Research Reagents and Tools

Item Name	Type	Function in Research	Example from Literature
Time-Lapse Incubator	Hardware	Provides the continuous imaging necessary to capture embryo morphokinetics, the primary data source for many models.	EmbryoScope+/EmbryoScope (Vitrolife) [61] [31] [17]
Preimplantation Genetic Testing for Aneuploidy (PGT-A)	Assay / Gold Standard	Provides the ground-truth labels (euploid/aneuploid) for training and validating supervised learning models.	Used across all cited ploidy prediction studies [8] [31] [17]
Vision Transformer (ViT) Masked Autoencoder	AI Architecture	A self-supervised learning framework used for pre-training foundation models on large volumes of unlabeled image data.	Used as the backbone for the FEMI model [31]
Bidirectional LSTM (BiLSTM)	AI Architecture	A type of recurrent neural network effective for analyzing sequential data, such as time-lapse video, to capture temporal dependencies.	Used in the BELA model for predicting blastocyst scores from video sequences [17]
WeightedROC Analysis	Statistical Method	A technique for standardizing performance metrics like AUC to account for differing covariate distributions (e.g., maternal age) across populations.	Used to mitigate the effect of age distribution differences between clinics [61]
SHapley Additive exPlanations (SHAP)	Software Library	Provides interpretability for AI models by quantifying the contribution of each input feature (e.g., specific time points) to the final prediction.	Used to analyze the importance of different development time points in the BELA model [17]

The pursuit of generalizable AI models for embryo ploidy prediction is a central challenge in reproductive medicine. While models like FEMI and BELA demonstrate promising performance (AUC ~0.76), evidence consistently shows that this performance can significantly degrade in external, multi-clinic validations [17]. Key factors impacting generalizability include varying patient demographics, particularly maternal age, and differing clinic-specific protocols [61].

To advance the field, the research community must prioritize methodologies that directly address these challenges. This includes the rigorous application of external validation, the adoption of statistical techniques like age-standardization for fair performance comparisons, and in-depth analysis of model stability and rank-order consistency [61] [58]. Future research should be directed toward developing more stable AI frameworks, leveraging larger and more diverse multicenter datasets for training, and ultimately, validating these tools based on clinically decisive endpoints such as live birth rates across diverse populations.

In the rapidly evolving field of artificial intelligence (AI) applications for embryo ploidy prediction, computational efficiency has emerged as a critical factor for successful clinical implementation. While numerous deep learning models demonstrate promising predictive capabilities for embryo euploidy, their real-world utility depends on effectively balancing model complexity with seamless integration into existing clinical workflows [36]. Embryo assessment represents a pivotal yet challenging step in in vitro fertilization (IVF), with conventional methods facing limitations including subjectivity, inter-observer variability, and labor-intensive processes [36].

The emergence of AI technologies, particularly deep learning algorithms using time-lapse imaging (TLI) data, offers promising solutions for automating embryo assessment and potentially increasing IVF success rates [63] [36]. However, these computational models vary significantly in their architectural complexity, data requirements, and computational demands, creating distinct trade-offs between predictive performance and practical implementation in diverse clinical settings. This comparative analysis examines current embryo ploidy prediction models through the critical lens of computational efficiency, evaluating how different architectural approaches balance sophisticated predictive capabilities with the practical constraints of clinical workflow integration.

Comparative Analysis of Model Architectures and Performance

Embryo ploidy prediction models employ diverse architectural approaches with varying computational requirements and performance characteristics. The table below summarizes key models, their architectures, and validated performance metrics:

Table 1: Comparative Analysis of Embryo Ploidy Prediction Models

Model Name	Architecture	Input Data	Performance (AUC)	Computational Requirements
BELA [17]	Multitask BiLSTM with ResNet backbone	Time-lapse videos (96-112 hpi) + maternal age	0.76 (EUP vs. ANU)	High (video processing, multiple focal planes)
iDAScore v2.0 [8]	Deep learning CNN	Time-lapse videos	0.68 (euploidy prediction)	Medium (integrated with EmbryoScope+ incubator)
STORK-A [17]	CNN	Single image (110 hpi)	~0.74 (literature reference)	Low (single image processing)
ERICA [17]	Deep learning CNN	Static embryo images	0.74	Low (static image analysis)
Random Forest Classifier [64]	Ensemble machine learning	Morphokinetic + clinical features	0.75	Low to medium (feature engineering dependent)

The integration of maternal age with time-lapse imaging data in the BELA model demonstrates how hybrid approaches can enhance performance without dramatically increasing computational complexity [17]. The model employs a two-step process where it first predicts a model-derived blastocyst score (MDBS) from processed day-5 time-lapse videos, then uses this score combined with maternal age to predict ploidy status through logistic regression [17]. This architectural decision represents a calculated balance between deep learning sophistication and practical predictive efficiency.

Conversely, the iDAScore system exemplifies clinical workflow integration through its direct compatibility with EmbryoScope+ incubators, providing real-time analysis without significant disruption to laboratory routines [8]. The system applies deep learning algorithms to time-lapse videos, assigning scores from 1.0 to 9.9 based on developmental patterns, and operates within existing clinical hardware infrastructure [8]. This integration strategy significantly reduces the computational overhead for clinics already utilizing Vitrolife's ecosystem.

Experimental Protocols and Methodologies

BELA Model Development and Validation

The BELA (Blastocyst Evaluation Learning Algorithm) framework employs a structured multitask learning approach optimized for ploidy prediction [17]:

Figure 1: BELA Model Workflow

Data Processing Pipeline: BELA processes time-lapse sequences typically comprising 360-420 distinct frames captured at 0.3-hour intervals over 5 days of development. The model specifically focuses on the blastocyst stage (96-112 hours post-insemination) based on ablation analyses comparing embryonic development time points [17].

Architecture Details: The model uses a pre-trained spatial feature extraction model to transform input videos into feature vectors. A multitasking Bidirectional LSTM (BiLSTM) model concurrently predicts inner cell mass (ICM), trophectoderm (TE), expansion, and blastocyst score components [17].

Training Methodology: Researchers trained and evaluated BELA using four-fold cross-validation on datasets from Weill Cornell Medicine's Center for Reproductive Medicine. The training incorporated 1998 Embryoscope time-lapse sequences and 841 sequences from Embryoscope+ systems [17].

iDAScore Clinical Validation

The iDAScore validation followed a comprehensive multi-center approach to assess real-world performance [8]:

Figure 2: iDAScore Validation Protocol

Validation Framework: Six retrospective studies meeting inclusion criteria formed the validation foundation, with all reporting statistically significant associations between higher iDAScore values and embryo euploidy. AUC values for euploidy prediction ranged from 0.60 to 0.68 across different studies and patient populations [8].

Integration Methodology: The iDAScore system was designed for direct integration into EmbryoScope+ incubators, allowing automatic analysis without requiring additional embryologist time or significant workflow modifications. This integration strategy represents a conscious design decision prioritizing computational efficiency and clinical practicality [8].

Computational Efficiency Analysis

Performance Versus Complexity Trade-offs

The evolution of embryo ploidy prediction models reveals distinct architectural strategies for balancing computational complexity with clinical utility:

Table 2: Computational Efficiency Comparison

Model Type	Inference Speed	Hardware Requirements	Clinical Scalability	Implementation Complexity
Video-based (BELA) [17]	Lower (video processing)	High (GPU acceleration)	Moderate (specialized hardware)	High (complex architecture)
Image-based (STORK-A) [17]	High (single image)	Low (CPU sufficient)	High (minimal infrastructure)	Low (streamlined processing)
Integrated (iDAScore) [8]	Medium (optimized hardware)	Medium (proprietary system)	Variable (vendor dependent)	Low (pre-integrated solution)
Feature-based (Random Forest) [64]	High (pre-computed features)	Low (standard computing)	High (flexible deployment)	Medium (feature engineering)

Video-based models like BELA demonstrate higher predictive accuracy (AUC 0.76) but require substantially greater computational resources for processing hundreds of time-lapse frames across multiple focal planes [17]. In contrast, image-based approaches like STORK-A offer faster inference times and lower hardware requirements while maintaining respectable performance (AUC ~0.74) [17].

The iDAScore system represents an intermediate approach, with performance (AUC 0.68) slightly below more complex models but with superior clinical workflow integration through its native implementation on EmbryoScope+ systems [8]. This architectural decision prioritizes operational efficiency and reproducibility across diverse clinical environments.

Clinical Workflow Integration Metrics

Successful clinical integration depends on multiple factors beyond raw predictive performance:

Processing Time Constraints: Models must provide predictions within clinical decision windows. Integrated systems like iDAScore generate scores in near real-time, while more complex models may require batch processing or cloud-based computation [8].

Interoperability Requirements: Compatibility with existing laboratory information management systems (LIMS) and electronic medical records (EMR) significantly impacts implementation complexity. Models requiring standalone interfaces or custom integration present higher adoption barriers [65].

Training and Expertise Demands: Systems that minimize the need for specialized technical expertise among clinical staff demonstrate higher adoption rates. The "black-box" nature of some complex deep learning models can create implementation resistance despite superior performance metrics [50].

Research Reagent Solutions

The experimental protocols for embryo ploidy prediction rely on specific research reagents and technical platforms that directly impact model performance and computational requirements:

Table 3: Essential Research Reagents and Platforms

Reagent/Platform	Function	Impact on Computational Efficiency
EmbryoScope+ System [8]	Time-lapse imaging with integrated analysis	Reduces external processing needs through native implementation
PicoPLEX Gold WGA Kit [66]	Whole genome amplification for PGT-A validation	Provides ground truth data for model training and validation
Takara Bio PicoPLEX Gold [66]	Single-cell DNA sequencing for ploidy confirmation	Enables high-quality training datasets for supervised learning
Vitrolife culture media [66]	Standardized embryo culture conditions	Reduces confounding variables in model development
NVIDIA T4 GPU [65]	Accelerated deep learning computation	Enables practical training times for complex video analysis models

Standardized reagent systems and platforms play a crucial role in computational efficiency by ensuring consistent input data quality and reducing preprocessing requirements. The use of commercial time-lapse systems with integrated AI capabilities represents a significant advancement toward computationally efficient clinical implementation [8] [65].

Discussion and Future Directions

The comparative analysis of computational efficiency in embryo ploidy prediction models reveals several critical considerations for clinical implementation. First, the trade-off between model complexity and practical utility necessitates careful evaluation of clinical context and available infrastructure. High-complexity models like BELA offer superior performance but require significant computational resources that may not be feasible in all clinical settings [17]. Second, integrated systems like iDAScore demonstrate how vendor-specific optimization can enhance workflow efficiency, though potentially at the cost of flexibility and interoperability [8].

Future research directions should focus on developing adaptive computational frameworks that can balance model complexity with available resources. Promising approaches include configurable architectures that can operate at different complexity levels based on clinical requirements, federated learning strategies to improve model generalization without centralized data aggregation, and hybrid systems that combine simpler rule-based algorithms with complex deep learning for specific edge cases [63] [36].

Additionally, the field would benefit from standardized computational efficiency metrics specific to clinical embryology applications, including processing time per embryo, hardware requirements, interoperability standards, and implementation complexity scores. Such metrics would enable more systematic comparisons across different architectural approaches and guide development of computationally efficient solutions that maintain predictive performance while enhancing clinical adoption [65] [36].

The evolution toward more computationally efficient embryo ploidy prediction will likely involve both technical innovations in model architecture and practical advances in clinical integration frameworks. By prioritizing computational efficiency alongside predictive accuracy, the field can develop solutions that deliver on the promise of AI-assisted embryo selection across diverse clinical settings and patient populations.

The selection of viable embryos is a cornerstone of successful in vitro fertilization (IVF). Preimplantation genetic testing for aneuploidy (PGT-A) serves as the gold standard for assessing embryonic ploidy status but is invasive, costly, and not universally applicable [8]. Consequently, artificial intelligence (AI) models have emerged as promising non-invasive alternatives for embryo evaluation. This guide provides a comparative analysis of contemporary AI models for embryo ploidy prediction, with a specific focus on their employed optimization techniques—namely multi-task learning and feature importance analysis. We objectively compare the performance of these models and detail the experimental protocols that validate their clinical utility for a research-oriented audience.

Comparative Performance of Embryo Ploidy Prediction Models

The performance of AI models in predicting embryo ploidy is quantitatively assessed using metrics such as the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). The following table summarizes the documented performance of various models, highlighting the impact of their underlying optimization techniques.

Table 1: Performance Comparison of Embryo Ploidy Prediction Models

Model Name	Core Optimization Technique	Key Input Data	Reported AUC for Ploidy Prediction	Key Performance Findings
BELA (Blastocyst Evaluation Learning Algorithm) [67] [17]	Multi-task Learning	Time-lapse sequences, Maternal age	0.76 (on Weill Cornell dataset) [67] [17]	Matches performance of models trained on embryologists' manual scores.
FEMI (Foundational IVF Model for Imaging) [57]	Self-Supervised Learning (Vision Transformer)	~18 million time-lapse images	Outperformed benchmark models (e.g., MoViNet, VGG16, EfficientNet) [57]	Superior accuracy in ploidy prediction, including under low embryo quality conditions.
LIFE Predict v1.1 [30]	Machine Learning (Ensemble Model)	Morphokinetic meta-variables, Clinical data	0.824 (Cross-validation), 0.818 (External Validation) [30]	Aneuploidy rates decreased across score quartiles (76.4% in lowest to 13.3% in highest).
Random Forest (XAI Model) [68]	Explainable AI (SHAP, LIME)	Morphokinetic features, Morphology grades, 11 clinical variables	0.808 (Internal), 0.750 (External Test Set) [68]	High accuracy; model decisions are interpretable.
iDAScore v2.0 [8]	Deep Learning (Convolutional Neural Network)	Time-lapse videos	AUC range: 0.60 - 0.68 (for euploidy prediction) [8]	Statistically significant association with euploidy; moderate predictive accuracy.
Gradient Boosting Model [49]	Image Processing (HOG + PCA)	Static embryo images	Accuracy: 0.74, Aneuploid Precision: 0.83 [49]	An efficient model using handcrafted image features.

Detailed Experimental Protocols and Model Architectures

Multi-Task Learning in the BELA Model

The Blastocyst Evaluation Learning Algorithm (BELA) employs a multi-task learning architecture to predict ploidy status. This approach involves a two-step process that leverages intermediate tasks to enhance the primary prediction goal [67] [17].

Protocol and Workflow:

• Input and Feature Extraction: Day 5 time-lapse videos (96–112 hours post-insemination) are processed. A pre-trained spatial feature extraction model transforms the video frames into feature vectors [17].
• Multi-Task Prediction (Blastocyst Score): A Bidirectional Long Short-Term Memory (BiLSTM) network is used to concurrently predict multiple, related sub-tasks: the inner cell mass (ICM) score, trophectoderm (TE) score, expansion score, and the overall blastocyst score (BS). This step produces a Model-Derived Blastocyst Score (MDBS) [17].
• Ploidy Prediction: The MDBS, along with maternal age as a continuous input feature, is fed into a logistic regression classifier to make the final euploid versus aneuploid prediction [17]. This architecture allows the model to learn generalized features from related morphological tasks, which improves its performance on the primary ploidy prediction task.

The following diagram illustrates the workflow and logical relationships within the BELA model:

BELA Model Multi-Task Workflow

Feature Importance Analysis in Explainable AI (XAI) Models

For models that function as "black boxes," Explainable AI (XAI) techniques are critical for interpreting predictions and building clinical trust. These techniques identify which input features most significantly impact the model's decision.

Protocol and Workflow: A study by Luong et al. utilized six different machine learning models, with Random Forest (RF) performing best for ploidy prediction (AUC: 0.808) [68]. To interpret this model, the researchers applied two XAI techniques:

• SHapley Additive exPlanations (SHAP): This global interpretation method quantifies the marginal contribution of each feature to the model's output across the entire dataset. The study identified maternal age, paternal age, time to blastocyst (tB), and day 5 morphology grade as the most impactful features for ploidy prediction [68].
• Local Interpretable Model-agnostic Explanations (LIME): This technique provides local explanations for individual predictions. It reveals how the model assigned a specific ploidy probability for a single embryo by showing the contribution value of each variable within a finite range for that particular case [68].

The application of these XAI techniques transforms an opaque prediction into an interpretable decision, providing researchers and clinicians with transparent insights into the factors driving the ploidy assessment.

The following diagram illustrates the process of explaining a ploidy prediction model using XAI:

XAI for Ploidy Model Interpretation

The Scientist's Toolkit: Essential Research Reagents and Materials

The development and validation of the AI models discussed rely on a foundation of specific biological materials, instrumentation, and software. The following table details key components of the experimental toolkit referenced in the studies.

Table 2: Key Research Reagent Solutions for Embryo Ploidy Prediction Studies

Item Name	Specific Type/Model	Function in the Research Context
Time-Lapse Incubator [8] [17] [49]	EmbryoScope+/EmbryoScope (Vitrolife), MIRI (Esco Medical)	Provides a stable culture environment while capturing continuous time-lapse imaging of embryo development, which is the primary data source for most AI models.
Inverted Microscope [49]	Olympus IX71, Nikon Eclipse Ti	Used for direct high-quality static image capture of embryos for models that utilize static images instead of videos.
Genetic Analysis Kit [49]	Veriseq PGS (Illumina)	Used for Next-Generation Sequencing (NGS) in PGT-A to determine the ground truth ploidy status of embryos for model training and validation.
Whole-Genome Amplification System [49]	SurePlex DNA Amplification System (Illumina)	Amplifies the DNA from biopsied trophectoderm cells to enable comprehensive genetic analysis via PGT-A.
Data Analysis Software [68] [49]	BlueFuse Multi (Illumina), SHAP/LIME Python libraries	Software for interpreting genetic data (BlueFuse) and for implementing explainable AI techniques to interpret machine learning model predictions (SHAP/LIME).

Validation Frameworks and Comparative Performance Metrics

Robust validation is the cornerstone of developing reliable artificial intelligence (AI) and machine learning (ML) models for embryo ploidy prediction. In the high-stakes field of assisted reproductive technology (ART), where models aim to non-invasively identify embryos with the highest implantation potential, distinguishing genuinely predictive algorithms from those that are overfitted to specific datasets is paramount. Two advanced methodological approaches have emerged as best practices for this task: internal-external cross-validation and multi-center validation [69] [62]. These frameworks rigorously test model performance across diverse clinical environments, patient populations, and laboratory protocols, providing evidence of generalizability that is essential for clinical translation. This guide objectively compares these validation approaches, detailing their experimental protocols and performance outcomes as implemented in contemporary embryo ploidy prediction research.

Comparative Analysis of Validation Frameworks

The table below summarizes the core objectives, key implementation characteristics, and representative performance outcomes associated with internal-external and multi-center validation approaches as applied in recent studies.

Table 1: Comparison of Internal-External and Multi-Center Validation Approaches

Feature	Internal-External Cross-Validation	Multi-Center External Validation
Core Objective	To simulate external validation using a series of internal hold-out tests, progressively validating on data from different clinics [69].	To assess model performance on a completely independent, unseen dataset collected from multiple external clinics [17] [30].
Key Implementation	Iteratively trains on data from (N-1) clinics and validates on the remaining one clinic, rotating until all clinics have served as the validation set [69].	Trains a model on one or more datasets and then tests it on a separate, independent dataset from one or multiple clinics not involved in training [17] [30].
Representative Performance	Logistic regression model: AUC 0.71 (95% CI 0.67-0.73) for ploidy prediction [69] [44].	BELA model: AUC 0.76 on external WCM-Embryoscope+ dataset [17]. LIFE Predict v1.1: AUC 0.818 in external validation [30].
Primary Advantage	Maximizes data usage for both training and validation while providing a robust estimate of performance across participating centers [69].	Provides the strongest evidence of real-world generalizability by testing on fully independent clinical environments and patient populations [30].
Common Challenges	Performance can be variable across different held-out clinics, reflecting site-specific biases [70].	Requires collaboration with external clinics and can be challenging due to data heterogeneity and protocol differences [17].

Detailed Experimental Protocols

Internal-External Cross-Validation Protocol

The internal-external cross-validation approach was rigorously implemented in a large-scale study comparing 12 machine learning models for ploidy prediction, which serves as a canonical protocol for this method [69] [44].

1. Data Pooling and Preparation: The study aggregated a meta-dataset of 8,147 biopsied blastocysts from 1,725 patients across nine IVF clinics in the UK [69] [44]. Each embryo was cultured in a time-lapse system, and the dataset included 22-26 covariates, including morphokinetic timings and clinical bio-data.

2. Iterative Validation Cycle: The process involved systematically rotating the validation set among all participating centers:

• For each iteration, data from N-1 clinics was used to train the model.
• The remaining one clinic was held out as the validation set.
• This process was repeated until every clinic had served as the validation set once [69].

3. Performance Aggregation: The performance metrics (e.g., AUC, F1-score) from each iteration were aggregated to produce a final estimate of model performance and its variability across clinical settings. The best-performing model in this framework was a mixed-effects logistic regression, which achieved an AUC of 0.71 and was notably superior to more complex machine learning models like random forest (AUC 0.68) and deep learning (AUC 0.63) approaches [69] [44].

Multi-Center External Validation Protocol

The multi-center external validation protocol is exemplified by the validation strategies of the BELA and LIFE Predict v1.1 models [17] [30].

1. Model Development on Internal Data:

• The BELA model was developed using data from Weill Cornell Medicine (WCM), comprising 1,998 Embryoscope time-lapse sequences [17].
• The LIFE Predict v1.1 model was trained on a dataset of 833 embryos from a consortium of Spanish fertility clinics [30].

2. Validation on Fully External Datasets:

• The trained models were then validated on completely separate datasets not involved in the training process.
• BELA was tested on an internal-external dataset (WCM-Embryoscope+ with 841 sequences) and a truly external dataset from IVI Valencia, Spain [17].
• LIFE Predict v1.1 was validated on a hold-out set of 357 embryos from different clinics within the ANACER network [30].

3. Performance Benchmarking: Model performance on the external validation sets was benchmarked against the development set performance and, in some cases, against other models or clinical standards. For instance, BELA's performance increased from an AUC of 0.66 to 0.76 when maternal age was included as an input feature during external validation [17].

Table 2: Performance of Models Undergoing Multi-Center External Validation

Model	Training Data	External Validation Data	Performance (AUC)
BELA [17]	1,998 embryos (WCM)	841 embryos (WCM-Embryoscope+)	0.76
LIFE Predict v1.1 [30]	833 embryos (ANACER clinics)	357 embryos (ANACER clinics)	0.818
iDAScore (across clinics) [70]	Internal test set	4,805 embryos (4 external clinics)	0.58 - 0.69 (Clinic-specific range)

Visualization of Validation Workflows

The following diagrams illustrate the logical structures and data flows for the two cross-validation approaches.

Internal-External Cross-Validation Workflow

Multi-Center External Validation Workflow

The Scientist's Toolkit: Research Reagent Solutions

The following table details key materials and computational tools essential for conducting rigorous validation studies in embryo ploidy prediction research.

Table 3: Essential Research Reagents and Tools for Ploidy Prediction Validation Studies

Reagent / Tool	Function / Application	Example Use in Research
Time-Lapse Incubator Systems	Provides stable culture conditions while capturing sequential embryo images for morphokinetic analysis [8] [36].	EmbryoScope+ system used to generate time-lapse videos for iDAScore analysis [8].
Preimplantation Genetic Testing for Aneuploidy (PGT-A)	Gold standard for establishing embryo ploidy ground truth; essential for model training and validation [69] [17].	Used as outcome label in 8,147 embryos to train and validate 12 machine learning models [69].
Deep Learning Frameworks (e.g., CNNs, BiLSTM)	Model architectures for processing image and time-series data to predict ploidy from time-lapse videos [17] [36].	BELA model uses BiLSTM to predict blastocyst score from day-5 time-lapse videos [17].
Statistical Software (R, Python)	Platforms for implementing cross-validation, performing statistical analysis, and calculating performance metrics [70].	Used for age-standardization of AUCs and weighted ROC analysis in multi-clinic comparisons [70].
Cloud-Based Data Platforms	Secure, centralized data storage and sharing for multi-center studies, enabling collaboration and external validation [30].	ANACLOUD platform used for safe data aggregation from nine Spanish fertility clinics [30].

Internal-external cross-validation and multi-center external validation are complementary, robust frameworks essential for developing clinically relevant embryo ploidy prediction models. The internal-external approach provides an efficient, resource-conscious method for obtaining realistic performance estimates during model development, as demonstrated by the large-scale comparison of 12 models [69]. In contrast, multi-center external validation represents the definitive test of model generalizability, with studies like BELA and LIFE Predict v1.1 showing that maintaining strong performance (AUC > 0.75) on completely independent datasets is achievable [17] [30]. For researchers, the choice between these methods is not binary; a rigorous validation strategy should ideally incorporate both, beginning with internal-external validation during development and culminating in multi-center external validation before clinical deployment. As the field progresses, standardization of these validation protocols will be crucial for objectively comparing models and ultimately translating the most reliable AI tools into IVF practice to improve patient outcomes.

Embryo ploidy status, indicating whether an embryo is chromosomally normal (euploid) or abnormal (aneuploid), is a critical determinant of successful implantation and live birth in in vitro fertilization (IVF). The selection of euploid embryos significantly enhances the likelihood of a successful pregnancy while reducing the risk of miscarriage [8] [9]. Traditionally, ploidy assessment has relied on preimplantation genetic testing for aneuploidy (PGT-A), an invasive and costly procedure that involves biopsy of trophectoderm cells [8]. This invasiveness has motivated the development of non-invasive assessment methods using artificial intelligence (AI) and machine learning (ML) algorithms that analyze time-lapse imaging and morphokinetic data.

Machine learning models offer a promising alternative by leveraging patterns in embryonic development to predict ploidy status without physical intervention. These models analyze extensive datasets of embryo images and morphokinetic parameters, capturing subtle developmental features associated with chromosomal normality [8] [9]. Performance benchmarking of these models is essential for clinical adoption, with the Area Under the Receiver Operating Characteristic Curve (AUC) serving as a key metric for evaluating predictive accuracy. This review provides a comprehensive comparative analysis of AUC performance across twelve machine learning models developed for embryo ploidy prediction, examining their methodological approaches, validation strategies, and clinical applicability.

Comparative Performance Analysis

Quantitative AUC Benchmarking Across Models

Table 1: AUC Performance Benchmarking of Embryo Ploidy Prediction Models

Model Name	AUC for Ploidy Prediction	Dataset Size (Embryos)	Key Predictors	Study Type
LIFE Predict v1.1	0.824 (internal); 0.818 (external)	1,190	Morphokinetic meta-variables, clinical data	Multicenter retrospective [30]
FEMI (Foundational Model)	>0.75	~18 million images	Time-lapse sequences, maternal age	Retrospective [9]
Mixed Effects Logistic Regression	0.71 (95% CI: 0.67-0.73)	8,147	Morphokinetic parameters, blastocyst expansion, trophectoderm grade	Multicenter cohort [69]
Random Forest Classifier	0.68	8,147	Morphokinetic parameters	Multicenter cohort [69]
iDAScore v2.0	0.68	249,635	Time-lapse morphokinetics	Retrospective multicentric [8]
Extreme Gradient Boosting	0.63	8,147	Morphokinetic parameters	Multicenter cohort [69]
Deep Learning Model	0.63	8,147	Morphokinetic parameters	Multicenter cohort [69]
iDAScore v1.0	0.60-0.67	3,448-3,604	Time-lapse morphokinetics	Multiple retrospective studies [8]
Oocyte Ploidy AI Model	0.66	177 blastocysts	Oocyte images, blastocyst development score	Retrospective [71]
Fused Clinical+Image AI Model	0.91 (clinical pregnancy)	1,503 cycles	Blastocyst images, clinical data (age, BMI)	International multicenter [72]
Random Forest (Live Birth)	>0.80	11,728 records	Female age, embryo grade, usable embryos, endometrial thickness	Retrospective [73]

Table 2: Performance Comparison by Algorithm Class

Algorithm Class	Best Performing Model	AUC Range	Key Advantages	Clinical Implementation Readiness
Ensemble Models	LIFE Predict v1.1	0.818-0.824	Integrates novel meta-variables with clinical data	High (externally validated) [30]
Traditional Statistical	Mixed Effects Logistic Regression	0.71	Handles clustered data, interpretable coefficients	Medium [69]
Deep Learning	FEMI Foundation Model	>0.75	Processes raw images, minimal manual annotation	Medium (computationally intensive) [9]
Tree-Based	Random Forest	0.68->0.80	Handles non-linear relationships, feature importance	Medium to High [69] [73]

Key Performance Insights

The benchmarking data reveals several critical patterns in model performance. First, the LIFE Predict v1.1 ensemble model demonstrated superior performance with AUC values of 0.824 in internal validation and 0.818 in external validation [30]. This model uniquely incorporates novel morphokinetic meta-variables (Range and MAEkinetic) that quantify deviations from normative development patterns observed in embryos that resulted in live births.

Second, foundation models like FEMI represent a significant advancement by leveraging self-supervised learning on massive image datasets (approximately 18 million time-lapse images) [9]. This approach achieves robust performance (AUC >0.75) while requiring minimal manual annotation of embryo characteristics.

Third, the comparative analysis of 12 models by Bamford et al. revealed that traditional statistical approaches (mixed effects logistic regression) can outperform more complex machine learning methods for ploidy prediction, achieving an AUC of 0.71 compared to 0.63-0.68 for other models [69]. This suggests that methodological sophistication does not always guarantee superior performance for this specific prediction task.

Finally, models that integrate multiple data types consistently outperform single-modality approaches. The fused clinical and image AI model achieved an exceptional AUC of 0.91 for clinical pregnancy prediction by combining blastocyst images with patient clinical information [72].

Experimental Protocols and Methodologies

Model Development Workflow

The following diagram illustrates the generalized experimental workflow for developing and validating embryo ploidy prediction models, synthesized from multiple studies:

Detailed Methodological Approaches

LIFE Predict v1.1 Development

The top-performing LIFE Predict v1.1 model employed a rigorous development methodology [30]. The retrospective multicenter study utilized data from 1,190 blastocysts across nine Spanish fertility clinics collected between 2017-2024. The model integrated clinical data with novel morphokinetic meta-variables:

• Range: Quantified the deviation of an embryo's morphokinetic timings from the average values observed in embryos that resulted in live births.
• MAEkinetic: Calculated the mean absolute error between an embryo's development pattern and the optimal trajectory.

The dataset was partitioned with 70% (n=833) for model training/testing and 30% (n=357) for external validation. The ensemble model architecture combined multiple algorithm types, with performance assessed via AUC-ROC and confusion matrix metrics. Logistic regression calculated odds ratios for aneuploidy risk across score quartiles.

FEMI Foundation Model Training

The FEMI (Foundational IVF Model for Imaging) approach represented a paradigm shift from task-specific models [9]. The methodology involved:

• Pre-training: A Vision Transformer masked autoencoder (ViT MAE) was pre-trained on approximately 18 million time-lapse embryo images using self-supervised learning.
• Architecture: The model utilized an encoder-decoder structure to learn domain-specific features from masked portions of embryo images.
• Data Processing: Images were cropped around embryos using a segmentation model based on InceptionV3 architecture, then resized to 224×224 pixels.
• Fine-tuning: The pre-trained encoder was subsequently fine-tuned for specific downstream tasks including ploidy prediction, with maternal age incorporated as an additional feature.

This foundation model approach demonstrated the scalability of leveraging large-scale unlabeled data to improve predictive accuracy across multiple embryology tasks.

12-Model Comparison Study

Bamford et al. conducted a comprehensive comparison of 12 machine learning models using a morphokinetic meta-dataset of 8,147 embryos [69]. The methodological framework included:

• Dataset: Multicenter cohort data from nine IVF clinics in the UK, including 3,004 euploid and 5,023 aneuploid embryos.
• Model Classes: Four different algorithmic approaches with three models each:
  • Mixed effects multivariable logistic regression
  • Random forest classifiers
  • Extreme gradient boosting
  • Deep learning
• Validation: Internal-external cross-validation and external validation procedures.
• Covariates: Two dataset configurations with 22 and 26 covariates respectively, including morphokinetic timings, blastocyst expansion, and trophectoderm grade.

This systematic comparison provided unique insights into the relative performance of different algorithmic approaches for ploidy prediction using consistent validation methodology.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Research Reagent Solutions for Embryo Ploidy Prediction Studies

Reagent/Technology	Function	Example Implementation
Time-Lapse Incubators	Continuous embryo monitoring without culture disturbance	EmbryoScope+ (Vitrolife) [8] [74]
Global Culture Media	Supports embryo development from fertilization to blastocyst	G-TL (Vitrolife) [74]
PGT-A Kits	Gold standard validation of ploidy status	Preimplantation genetic testing for aneuploidy kits [8] [30]
Vitrification Systems	Embryo cryopreservation for subsequent transfer	Closed CBS High Security Vitrification straws [74]
Image Analysis Software	Automated morphokinetic parameter annotation	EmbryoViewer software (Vitrolife) [74]
Hormone Assays	Assessment of ovarian reserve and cycle monitoring	Anti-Müllerian hormone (AMH), estradiol (E2), progesterone (P4) tests [73] [72]

This comprehensive benchmarking analysis reveals significant variability in the performance of machine learning models for embryo ploidy prediction, with AUC values ranging from 0.60 to 0.824 across different algorithmic approaches. The superior performance of ensemble models like LIFE Predict v1.1 and foundation models like FEMI highlights the importance of integrating multiple data types and leveraging large-scale training datasets. However, the strong showing of traditional statistical methods like mixed effects logistic regression reminds us that algorithmic complexity alone does not guarantee predictive superiority.

For researchers and clinicians, these findings suggest that the optimal model choice depends on specific clinical requirements, available data types, and implementation constraints. While models with higher AUC values generally offer better discriminative ability, factors such as interpretability, computational requirements, and validation robustness should also inform selection decisions. Future research directions should prioritize prospective validation studies, standardization of performance metrics across clinics, and development of more sophisticated ensemble approaches that leverage the complementary strengths of different algorithmic families.

Within the realm of assisted reproductive technology, the selection of embryos with the correct number of chromosomes, known as euploidy, is a critical determinant of successful implantation and live birth. The comparative analysis of models for predicting embryo ploidy centers on a fundamental divide: invasive biopsy-based methods versus emerging non-invasive artificial intelligence (AI) techniques. For researchers and drug development professionals, understanding this landscape is crucial for directing future research, allocating resources, and developing next-generation diagnostic platforms.

This guide provides an objective comparison of the diagnostic accuracy and clinical utility of these competing paradigms. It synthesizes current experimental data and details the essential methodologies and reagents that form the foundation of this rapidly evolving field.

Methodological Comparison: Core Technologies and Workflows

The invasive and non-invasive approaches for embryo ploidy prediction are fundamentally different in their execution, from initial handling to final genetic analysis.

Invasive Method: Preimplantation Genetic Testing for Aneuploidy (PGT-A)

Preimplantation Genetic Testing for Aneuploidy (PGT-A) is the established invasive method for determining embryonic ploidy status. It involves a physical biopsy of cells from the blastocyst-stage embryo [75] [76].

• Experimental Protocol: The standard PGT-A workflow is a multi-step process. On day 5 or 6 post-fertilization, a laser is used to create an opening in the zona pellucida. Subsequently, multiple cells from the trophectoderm (TE), the precursor to the placenta, are aspirated via a biopsy micropipette [75]. The biopsied cells are then subjected to genetic analysis, typically using comprehensive chromosome screening (CCS) methods like next-generation sequencing (NGS) to quantify chromosomal copy numbers. The remaining embryo is cryopreserved while the genetic analysis is completed, with transfer occurring in a subsequent cycle.

• Limitations and Risks: As a biopsy-based method, PGT-A is inherently invasive. The procedure requires specialized equipment and highly trained embryologists, is time-consuming, and adds significant cost to the IVF process [77] [75] [76]. More critically, the biopsy process itself raises concerns about potential harm to the embryo's developmental potential. Furthermore, some studies associate it with increased risks of certain obstetric complications, such as preeclampsia and placenta previa [76].

Non-Invasive Method: Artificial Intelligence (AI) Analysis

Non-invasive ploidy prediction leverages artificial intelligence (AI) to assess embryo health without a biopsy. These models analyze data such as microscopic images of the embryo to predict the likelihood of euploidy.

• Experimental Protocol: The workflow for non-invasive AI prediction is significantly more streamlined. Embryos are cultured in a time-lapse imaging system that automatically captures thousands of images throughout the first five days of development [78]. Key features are then extracted from this image data. Different AI models utilize different inputs:

  • Morphokinetic Parameters: These are precise timings of key developmental events, such as the time of second polar body extrusion (tPB2), pronuclei appearance (PN), and the time to reach the 7-cell stage (t7) [79].
  • Morphological Features from 3D Modeling: Advanced image analysis can construct 3D models of the blastocyst to measure parameters like blastocyst diameter, TE cell number, TE cell density, and the area of the inner cell mass (ICM) [75].
  • The extracted features are fed into a trained AI model, such as a deep learning algorithm (e.g., convolutional neural networks) or a machine learning classifier (e.g., Random Forest, Gradient Boosting) [78] [79]. The model then outputs a prediction of euploidy or aneuploidy.

• Advantages: The primary advantage is the complete absence of embryo manipulation required for a biopsy, eliminating any associated risks to the embryo. It is also faster and less expensive than PGT-A [78] [76].

The following workflow diagrams illustrate the key steps for each of these core methods.

Workflow Comparison: PGT-A vs. AI Analysis

Diagnostic Performance: A Quantitative Comparison

The most critical metric for comparing these methods is their diagnostic accuracy, as measured by sensitivity, specificity, and area under the curve (AUC) in predicting embryonic euploidy.

Table 1: Diagnostic Accuracy of Ploidy Prediction Methods

Method	Representative Model/Technique	Sensitivity	Specificity	AUC	Key Findings
Invasive (PGT-A)	Trophectoderm Biopsy + NGS	Gold Standard	Gold Standard	N/A	Considered the diagnostic reference; provides direct genetic information but is invasive.
Non-Invasive AI	STORK-A Algorithm [77]	~70% (Overall)	~70% (Overall)	N/A	Accuracy for predicting non-euploidy; accuracy for complex aneuploidy: 77.6%.
Non-Invasive AI	BELA Algorithm [78]	N/A	N/A	0.82	Deep learning model using time-lapse imaging and孕龄.
Non-Invasive AI	Decision Tree Model [75]	96.2%	94.7%	0.978	Model based on 3D blastocyst parameters (e.g., TE cell number, ICM area).
Non-Invasive AI	Meta-Analysis (2024) [76]	0.67 (Pooled)	0.58 (Pooled)	0.67	Systematic review of 20 studies. Performance improved with top models.
Non-Invasive AI	Meta-Analysis (Top Models) [76]	0.71 (Pooled)	0.75 (Pooled)	0.80	Analysis restricted to the highest-accuracy model from each study.

The data reveals a performance spectrum for non-invasive AI. While a large-scale meta-analysis indicates modest aggregate performance (sensitivity 0.67, specificity 0.58) [76], specific, optimized models demonstrate that high accuracy is feasible. For instance, one model using 3D morphological parameters reported exceptional sensitivity (96.2%) and specificity (94.7%) [75]. Furthermore, AI models that integrate morphokinetic features—such as the timing of cell divisions—with clinical data like maternal age tend to perform better than those relying on images alone [79] [76].

The Scientist's Toolkit: Essential Research Reagents and Materials

The development and implementation of these ploidy prediction models rely on a suite of specialized reagents and platforms. The following table details key materials for researchers in this field.

Table 2: Key Research Reagent Solutions for Embryo Ploidy Research

Item	Function in Research	Specific Examples / Context
Time-Lapse Incubators	Provides a stable culture environment while automatically capturing sequential images of embryo development for morphokinetic analysis.	Used in training AI models like BELA [78].
Biopsy Micropipettes	Essential for performing the invasive TE biopsy for PGT-A; used to aspirate cells from the embryo.	A critical tool for the gold-standard method and for creating labeled datasets to train AI models [75].
Next-Generation Sequencing (NGS) Kits	For comprehensive chromosome analysis of biopsied cells in PGT-A. Provides the "ground truth" ploidy status.	Used to generate validated datasets for training and testing non-invasive AI algorithms [75] [76].
AI/ML Software Frameworks	Platforms for developing and training machine learning and deep learning models on embryo image and data sets.	Convolutional Neural Networks (CNNs) [75], Random Forest Classifiers (RFC), Gradient Boosting (GB) machines [79].
High-Performance Computing (HPC)	Provides the computational power required for training complex AI models on large datasets of embryo images.	NVIDIA A40 GPUs used in the BioHPC cluster to train the BELA model [78].
Image Analysis Software	Used for segmenting embryo components, constructing 3D models, and quantifying morphological parameters.	U-Net models for segmenting TE and ICM cells from images [75].

The current evidence does not support a wholesale replacement of invasive PGT-A by non-invasive AI. Instead, the future lies in a synergistic application of both methods to maximize clinical benefit and minimize risk. AI's most immediate and promising role is as a powerful triage tool. By pre-screening embryos and identifying those with a high probability of aneuploidy, AI can help clinicians decide which embryos warrant the cost and invasiveness of confirmatory PGT-A [78] [76]. This integrated protocol can make the IVF workflow more efficient and cost-effective.

For researchers and drug development professionals, the path forward is clear. Future work must focus on the external validation and standardization of AI models across diverse clinical settings and populations [76]. There is a significant need for large, multi-center, prospective studies to move these tools from research prototypes to clinically validated instruments. Furthermore, the exploration of multimodal AI, which combines time-lapse imaging, 3D morphology, and clinical biomarkers, holds the greatest potential to bridge the diagnostic accuracy gap with PGT-A, ultimately advancing the goal of achieving a single, healthy live birth.

Embryo selection remains a pivotal challenge in assisted reproductive technology (ART), with the ultimate goal of achieving a healthy, term live birth. The comparative analysis of embryo assessment technologies focuses on their predictive value for two critical clinical endpoints: live birth rates (LBR) and miscarriage rates. This review objectively compares the performance of preimplantation genetic testing for aneuploidy (PGT-A) against emerging artificial intelligence (AI)-based non-invasive models, evaluating their respective capacities to prognosticate these primary outcomes within the context of a broader thesis on comparative analysis of embryo ploidy prediction models research.

Comparative Performance of Embryo Assessment Modalities

Preimplantation Genetic Testing for Aneuploidy (PGT-A)

PGT-A, an invasive genetic tool, represents the current clinical standard for direct chromosomal assessment. However, recent high-quality evidence questions its efficacy for improving cumulative live birth rates (CLBR), particularly in specific patient populations.

Table 1: Clinical Outcomes for PGT-A vs. Conventional IVF/ICSI in RPL Patients

Outcome Measure	PGT-A Group	Conventional IVF/ICSI Group	Statistical Significance (P-value)	Study Details
Conservative CLBR (Cycle 1)	-	-	aOR=0.78 (95% CI: 0.49–1.23)	P > 0.05 [80]
Conservative CLBR (Cycle 3)	-	-	aOR=0.96 (95% CI: 0.60–1.53)	P > 0.05 [80]
Time to Live Birth	Significantly Longer	Shorter	aHR=0.56 (95% CI: 0.42–0.75)	P < 0.05 [80]
Miscarriage Rate	No significant difference	No significant difference	P > 0.05	Comparable [80]

A 2025 retrospective cohort study of Recurrent Pregnancy Loss (RPL) patients concluded that PGT-A did not significantly improve CLBR or shorten the time to live birth compared to conventional IVF/ICSI. The time to achieve a live birth was significantly prolonged in the PGT-A group, a critical consideration for patients and clinicians [80]. This aligns with a 2024 committee opinion from the American Society for Reproductive Medicine (ASRM), which states that the value of PGT-A as a routine screening test to lower miscarriage risk or improve live birth rates for all IVF patients has not been demonstrated [15].

Non-Invasive AI-Based Ploidy and Outcome Prediction Models

As alternatives to invasive biopsy, several AI and machine learning (ML) models have been developed to predict ploidy status and clinical outcomes non-invasively using time-lapse imaging and morphokinetic data.

Table 2: Performance of Non-Invasive AI/ML Embryo Assessment Models

Model Name	Primary Function	Key Finding Related to Live Birth/Miscarriage	Performance Metrics	Source
iDAScore (v1.0 & v2.0)	Deep learning-based embryo viability score	Higher scores positively associated with live birth; negatively associated with miscarriage [8].	AUC for euploidy prediction: 0.60-0.68 [8]
PREFER-MK Model	Morphokinetic-based aneuploidy risk categorization	"Low risk" embryos significantly more likely to result in live birth vs. "high risk" (OR=1.95, 95% CI:1.65–2.25). No significant association with miscarriage [81] [82].	Live Birth Rates: "High Risk": 38%, "Moderate Risk": 49%, "Low Risk": 50% [81] [82]
LIFE Predict v1.1	Machine learning model using morphokinetic meta-variables	Significant inverse relationship between model score and aneuploidy risk; stratifies live birth potential within morphological grades [30].	AUC: 0.818 (external validation); Aneuploidy rate in highest score quartile: 13.3% [30]
BELA Model	Automated ploidy prediction from time-lapse	Predicts ploidy status and blastocyst score without manual annotation, correlating with implantation potential [17].	AUC (EUP vs. ANU): 0.76 (with maternal age) [17]

These models demonstrate a consistent, moderate association between morphokinetic patterns and embryo ploidy or viability. The PREFER-MK model shows a clinically relevant doubling of the odds for live birth when comparing "low risk" to "high risk" embryos [81] [82]. Similarly, the LIFE Predict v1.1 model effectively stratifies embryos by risk, demonstrating that even within the same morphological grade, aneuploidy rates can vary dramatically from 11-14% (highest score quartiles) to 68-85% (lowest quartiles), directly impacting potential live birth outcomes [30].

Detailed Experimental Protocols

Protocol 1: PGT-A Clinical Outcome Study in RPL Patients

• Study Design: Retrospective cohort study.
• Participants: RPL patients (≥2 miscarriages) undergoing their first oocyte retrieval and at least one single-blastocyst transfer between June 2016 and June 2022. Patients with uterine anomalies, autoimmune diseases, or using donor sperm were excluded [80].
• Intervention Group (PGT-A): Blastocysts underwent trophectoderm biopsy. Genetic analysis was performed using comprehensive chromosome screening (e.g., NGS) to identify euploid embryos for transfer [80] [15].
• Control Group (Conventional IVF/ICSI): Embryos were selected for transfer based on standard morphological assessment without genetic testing [80].
• Primary Outcomes: Cumulative Live Birth Rate (CLBR) after up to three transfer cycles, live birth rate per transfer, miscarriage rate, and time to live birth [80].
• Statistical Analysis: Used adjusted odds ratios (aOR) and hazard ratios (aHR) with 95% confidence intervals (CI) to compare groups, controlling for confounders like maternal age and number of previous miscarriages [80].

Protocol 2: Development and Validation of AI/ML Prediction Models

• Data Acquisition: Retrospective collection of time-lapse imaging (TLI) videos from embryos cultured in integrated incubator systems (e.g., EmbryoScope+). datasets often include thousands of embryos with known outcomes (live birth) or known ploidy status (via PGT-A) [17] [30] [36].
• Data Annotation and Preprocessing: For models predicting ploidy, PGT-A results serve as the ground truth. Embryos are labeled as euploid (EUP) or aneuploid (ANU). For live birth prediction, embryos are labeled based on known transfer outcomes [17] [58]. TLI videos are processed into feature vectors.
• Model Architecture and Training:
  • Deep Learning (e.g., iDAScore, BELA): Utilizes Convolutional Neural Networks (CNNs) and sometimes Recurrent Neural Networks (RNNs) like BiLSTM to analyze image sequences and morphokinetic data [17] [36]. Models are trained to output a continuous score (e.g., 1.0-9.9) or a ploidy probability.
  • Machine Learning (e.g., LIFE Predict v1.1, PREFER): Integrates morphokinetic parameters (e.g., timings of cell divisions) and clinical data (e.g., maternal age) using algorithms like logistic regression, random forest, or gradient boosting [81] [30]. Some models create novel "meta-variables" quantifying developmental deviations from live birth-associated patterns [30].
• Validation: Performance is evaluated using held-out test sets or external validation cohorts from multiple fertility centers. Metrics include Area Under the Curve (AUC), accuracy, and odds ratios (OR) for live birth and miscarriage [8] [17] [30].

Workflow and Logical Relationship Diagram

The following diagram illustrates the logical relationship between embryo assessment technologies, their immediate predictions, and the ultimate clinical outcomes of live birth and miscarriage.

The Scientist's Toolkit: Essential Research Reagents and Materials

Table 3: Key Reagent Solutions for Embryo Ploidy Prediction Research

Item/Solution	Function in Research Context
Time-Lapse Incubator System	Provides a stable culture environment while continuously capturing images of embryo development, generating the essential video dataset for AI model training and validation [8] [36].
Trophectoderm Biopsy Kit	Enables the physical removal of cells from the blastocyst for PGT-A, establishing the genetic ground truth for model development and serving as the core intervention for PGT-A outcome studies [80] [15].
Next-Generation Sequencing (NGS) Kit	Performs comprehensive 24-chromosome analysis of biopsied samples, providing the high-resolution ploidy data used as a gold standard label for supervised learning of AI models [17] [15].
Annotation Software Platform	Allows embryologists to manually grade embryo morphology and annotate key morphokinetic timings, creating labeled datasets for traditional analysis and for training supervised AI algorithms [17] [30].
Pre-trained Convolutional Neural Network (CNN) Models	Serve as the foundational architecture for feature extraction from time-lapse images or videos, forming the backbone of deep learning-based assessment tools like iDAScore and BELA [17] [36].

The comparative analysis reveals a nuanced landscape. PGT-A, while directly assessing chromosomal content, has not conclusively demonstrated superior cumulative live birth rates compared to conventional methods in all patient populations, such as those with RPL, and may prolong the time to achieve a pregnancy [80]. In contrast, non-invasive AI/ML models show significant promise by providing a risk stratification that is associated with live birth outcomes, as evidenced by the PREFER-MK and LIFE Predict models [81] [30]. These tools can refine selection within morphologically similar embryos, potentially identifying hidden viability factors.

However, a critical consideration for AI models is stability and reliability. A 2025 study evaluating the stability of AI models for embryo selection found substantial inconsistency in embryo rank-ordering and high critical error rates among replicate models, raising concerns about their current readiness for unguided clinical deployment [58].

In conclusion, while PGT-A remains a valuable tool for specific indications, its universal application to improve live birth and reduce miscarriage rates is not strongly supported by recent evidence. Non-invasive AI models represent a powerful emerging adjunct, capable of associating developmental patterns with live birth potential. The future of embryo selection likely lies in integrated approaches, but requires robust, prospectively validated, and stable AI systems before they can be considered a new standard of care.

The integration of artificial intelligence (AI) into reproductive medicine has revolutionized embryo selection in in vitro fertilization (IVF), with deep learning models emerging as powerful tools for predicting embryo ploidy status. These models analyze time-lapse imaging and morphological data to non-invasively assess embryonic viability, offering a promising alternative to invasive preimplantation genetic testing for aneuploidy (PGT-A) [8] [17]. However, as these technologies advance, significant limitations persist in their ability to accurately detect specific types of chromosomal abnormalities, particularly segmental aneuploidies and complex ploidy anomalies.

Segmental aneuploidies—partial chromosomal gains or losses involving chromosome segments larger than 5 Mb—present a substantial challenge for current prediction models. These abnormalities occur in approximately 4.5-8.4% of blastocysts and originate from diverse mechanisms including chromothripsis, mitotic errors, or technical artifacts during biopsy and analysis [83]. Despite their clinical significance, AI models demonstrate substantially reduced performance in identifying these abnormalities compared to whole-chromosome aneuploidies, creating a critical gap in non-invasive embryo assessment capabilities.

This review provides a comprehensive analysis of the technical limitations underlying current ploidy prediction models, with particular focus on their performance disparities in detecting segmental versus whole-chromosome abnormalities. By examining experimental data, methodological constraints, and emerging solutions, we aim to inform researchers and clinicians about the current capabilities and limitations of these technologies in clinical practice.

Performance Disparities in Detecting Segmental vs. Whole-Chromosome Abnormalities

Quantitative Performance Metrics Across Abnormality Types

Current ploidy prediction models exhibit markedly different performance characteristics when detecting various types of chromosomal abnormalities. The following table summarizes the documented efficacy of leading models across abnormality categories:

Table 1: Performance Comparison of Ploidy Prediction Models Across Abnormality Types

Model/Approach	Abnormality Type	AUC	Sensitivity	Specificity	Clinical Context
iDAScore v1.0 [8]	Euploidy vs. Aneuploidy	0.60-0.68	N/A	N/A	Broad embryo screening
BELA [17]	Euploidy vs. Complex Aneuploidy	0.826	N/A	N/A	With maternal age integration
BELA [17]	Euploidy vs. All Aneuploidy	0.76	N/A	N/A	With maternal age integration
PGT-Plus AI Model [84]	Triploidy	1.00	100%	100%	Specialized ploidy detection
PGT-Plus AI Model [84]	Genome-Wide UPD	1.00	100%	100%	Specialized ploidy detection
TE Biopsy (PGT-A) [85]	Whole-Chromosome Aneuploidy	N/A	98.1%	100%	Invasive genetic testing
TE Biopsy (PGT-A) [85]	Segmental Aneuploidy	N/A	94.4%	38.7%	Invasive genetic testing

The performance disparity is particularly evident when comparing model efficacy for different abnormality types. While specialized AI models like PGT-Plus achieve perfect detection for triploidy and genome-wide uniparental diploidy (GW-UPD), general-purpose ploidy prediction models like iDAScore and BELA show more modest performance for comprehensive aneuploidy detection [8] [84]. This suggests that model architecture and training data specificity significantly impact detection capabilities for different abnormality categories.

Biological and Technical Foundations of Detection Limitations

The fundamental challenge in detecting segmental abnormalities stems from several biological and technical factors. Biologically, segmental aneuploidies affect only portions of chromosomes, potentially manifesting more subtle morphological phenotypes than whole-chromosome abnormalities. This reduces the discriminatory power of image-based AI models that rely on morphological and morphokinetic parameters [8] [17].

Technically, the limited concordance between trophectoderm (TE) and inner cell mass (ICM) in segmentally abnormal embryos compounds detection challenges. Research demonstrates that TE-ICM concordance rates are significantly lower for segmental aneuploidies (25%) compared to whole-chromosome aneuploidies (94%) or euploid embryos (85%) [85]. This biological discrepancy means that even accurate TE assessment may not reflect the true embryonic genotype, particularly for segmental abnormalities.

Table 2: Trophectoderm-Inner Cell Mass Concordance by Abnormality Type

Ploidy Status	TE-ICM Concordance Rate	ICM Euploidy Rate	Clinical Implications
Euploid	85%	85%	High confidence in transfer
Whole-Chromosome Aneuploidy	94%	0%	Reliable exclusion
Segmental Aneuploidy	25%	19%	Low prediction reliability
Segmental Mosaicism	33%	63%	Moderate prediction reliability

Additionally, the origin and characteristics of segmental abnormalities impact detectability. Segmental aneuploidies are more frequent in medium-sized metacentric or submetacentric chromosomes and particularly in q-chromosome arms [83]. Their size variation (typically >5Mb) and potential mosaic distribution further complicate consistent detection across different platforms and models.

Experimental Approaches and Methodological Limitations

Deep Learning Model Architectures and Training Paradigms

Current ploidy prediction models employ diverse architectural approaches with varying limitations for abnormality detection:

Time-lapse video analysis models like BELA (Blastocyst Evaluation Learning Algorithm) utilize multitask learning to predict blastocyst scores from day-5 time-lapse videos (96-112 hours post-insemination), then apply logistic regression with maternal age to predict ploidy status [17]. This approach achieves an AUC of 0.76 for euploid versus all aneuploid embryos and 0.826 for euploid versus complex aneuploid embryos when maternal age is incorporated. However, the model's performance depends heavily on blastocyst score prediction accuracy, with a mean absolute error of 1.855±0.03 compared to embryologist-assigned scores [17].

Integrated genetic analysis models like PGT-Plus employ ultra-low-coverage whole-genome sequencing (ulc-WGS) data and random forest algorithms to detect ploidy abnormalities, achieving near-perfect accuracy for triploidy and GW-UPD [84]. This method analyzes heterozygosity rates of high-frequency biallelic SNPs and likelihood ratios of alleles under different inheritance assumptions, leveraging allele frequencies and linkage disequilibrium from reference databases. While highly accurate for specific ploidy abnormalities, this approach requires genetic material and cannot be applied non-invasively.

Hybrid image-based deep learning models like iDAScore use convolutional neural networks (CNNs) trained on extensive time-lapse video datasets with known clinical outcomes, assigning scores from 1.0 to 9.9 based on developmental patterns [8]. These models demonstrate significant association with embryo euploidy (AUC 0.60-0.68) but show only moderate predictive accuracy when restricted to euploid embryo cohorts, suggesting limited detection capability for abnormalities that don't manifest morphologically [8].

Critical Reagents and Research Tools

Table 3: Essential Research Reagents and Platforms for Ploidy Detection Studies

Reagent/Platform	Function	Detection Limitations
Next-Generation Sequencing (NGS) [84] [86]	Comprehensive aneuploidy detection via low-pass whole-genome sequencing	Limited resolution for segments <5-10Mb; requires TE biopsy
Ion ReproSeq PGS Kit [83]	Whole genome amplification for PGT-A	Potential introduction of artifacts misinterpreted as segmental imbalances
EmbryoScope+/EmbryoScope [8] [17]	Time-lapse imaging for morphokinetic analysis	Limited phenotypic correlation with segmental abnormalities
SNP Microarrays [86]	Detection of subchromosomal anomalies via SNP profiling	Limited ability to detect small structural aberrations (<5Mb)
aCGH Platforms [86]	Genome-wide copy number variant detection	Cannot identify haploid/polyploid embryos or balanced rearrangements

The following diagram illustrates the experimental workflow and failure points in segmental aneuploidy detection:

Analytical Framework: Understanding Detection Failure Mechanisms

The limited ability of AI models to detect segmental aneuploidies stems from multiple biological and technical factors:

Biological Discordance: The low concordance (25%) between trophectoderm (TE) biopsy results and the actual inner cell mass (ICM) genotype in segmentally abnormal embryos represents a fundamental biological limitation [85]. This discrepancy means that even perfect biopsy analysis may not reflect true embryonic ploidy status. The ICM euploidy rate of 19% in embryos classified as segmentally aneuploid by TE biopsy further complicates prediction accuracy [85].

Technical Artifacts: Whole genome amplification (WGA)—required for PGT-A from limited biopsy material—introduces artifacts including allele drop-out, preferential amplification, and structural DNA anomalies that can be misinterpreted as segmental imbalances [83]. S-phase artifacts, where single-cell DNA replication domains result in copy number changes interpreted as segmental aneuploidy, present additional technical challenges [83].

Resolution Thresholds: Standard NGS-based PGT-A methodologies typically have detection thresholds of 5-10Mb for segmental abnormalities, potentially missing smaller but clinically significant segments [83] [86]. While increasing sequencing depth can improve resolution, practical and economic constraints limit implementation in clinical settings.

Morphological Correlation Gap: Segmental aneuploidies likely produce more subtle morphological phenotypes than whole-chromosome abnormalities, reducing the discriminatory power of image-based AI models [8] [17]. This phenotypic subtlety means that even advanced deep learning models analyzing time-lapse imaging may lack sufficient features for reliable detection.

Methodological Comparison and Performance Gaps

The following diagram illustrates the performance disparities across different detection methodologies:

Current ploidy prediction models demonstrate significant limitations in detecting segmental aneuploidies and specific ploidy abnormalities despite advancing capabilities in whole-chromosome aneuploidy detection. The performance disparity stems from biological factors like TE-ICM discordance, technical constraints including resolution thresholds and amplification artifacts, and methodological challenges in correlating morphological features with genetic abnormalities.

For researchers and clinicians, these limitations highlight the necessity of complementary approaches when segmental abnormalities are suspected. Specialized genetic analysis like the PGT-Plus model offers solutions for specific ploidy abnormalities but requires invasive biopsy [84]. Image-based AI models provide valuable non-invasive screening but cannot reliably exclude segmental anomalies [8] [17].

Future research directions should focus on integrating multi-modal data streams—combining time-lapse imaging with spent culture medium analysis or developing advanced algorithms specifically trained on segmental abnormality datasets. Additionally, improving the resolution of non-invasive genetic analysis from spent culture medium could potentially bridge current detection gaps without compromising embryo viability.

Understanding these model limitations is essential for proper clinical implementation and setting realistic expectations regarding the detection capabilities of current ploidy prediction technologies. As the field advances, acknowledging these constraints will guide the development of more comprehensive solutions for complete embryonic chromosomal assessment.

Conclusion

The comparative analysis reveals a rapidly evolving field where non-invasive AI models demonstrate promising but moderate predictive accuracy for embryo ploidy status, with AUC values typically ranging from 0.60-0.76. While these approaches cannot yet replace PGT-A as a standalone diagnostic, they offer valuable prioritization tools when genetic testing is not feasible and represent a paradigm shift toward less invasive embryo selection. Future directions should focus on prospective validation in diverse clinical settings, improved detection of mosaicism and segmental aneuploidies, and integration of multi-modal data sources. For biomedical research, the development of optimized algorithms that incorporate minimal-necessary covariates while maintaining clinical utility remains a critical challenge. The convergence of AI technology with embryology promises to enhance standardization, reduce costs, and potentially improve reproductive outcomes, though rigorous validation and consideration of ethical implications must guide clinical implementation.

References

• 1. AI-Driven Fluorescence Peak Analysis for Chromosomal ... [https://www.medrxiv.org/content/10.1101/2025.03.22.25324455v1.full-text]
• 2. Modeling and targeting general and chromosome-specific ... [https://genesdev.cshlp.org/content/early/2025/07/25/gad.352766.125]
• 3. Prevalence of chromosomal alterations in first-trimester ... [https://www.nature.com/articles/s41591-023-02645-5]
• 4. The impact of maternal age on aneuploidy in oocytes [https://www.sciencedirect.com/science/article/abs/pii/S1568163724001107]
• 5. Outcomes of pregnancies that screened positive for sex ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12435101/]
• 6. Fetal karyotyping in adolescent pregnancies: a population- ... [https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2025.1581249/full]
• 7. Performance and clinical implications of non-invasive ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12404953/]
• 8. Deep Learning Models: Their Relationship with Embryonic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12385821/]
• 9. A foundational model for in vitro fertilization trained on 18 ... [https://www.nature.com/articles/s41467-025-61116-2]
• 10. Preimplantation genetic testing: A narrative review - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11236403/]
• 11. Weighing The Pros And Cons of PGT-A [https://www.audubonfertility.com/post/pros-and-cons-of-pgt-a]
• 12. Preimplantation genetic testing for aneuploidy: challenges in ... [https://humgenomics.biomedcentral.com/articles/10.1186/s40246-022-00442-8]
• 13. Preimplantation Genetic Testing for Genetic Diseases [https://pmc.ncbi.nlm.nih.gov/articles/PMC10671162/]
• 14. A systematic review and meta-analysis of the diagnostic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12077728/]
• 15. The use of preimplantation genetic testing for aneuploidy [https://www.asrm.org/practice-guidance/practice-committee-documents/the-use-of-preimplantation-genetic-testing-for-aneuploidy-a-committee-opinion-2024/]
• 16. A systematic review and meta-analysis of the diagnostic ... [https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0321859]
• 17. Automatic ploidy prediction and quality assessment of ... [https://www.nature.com/articles/s41467-024-51823-7]
• 18. Fluorescence In Situ Hybridization (FISH) and Its Applications [https://pmc.ncbi.nlm.nih.gov/articles/PMC7122835/]
• 19. Comparison of array comparative genomic hybridization and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC4463508/]
• 20. Methods for comprehensive chromosome screening of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3348286/]
• 21. Array-based technology and recommendations for ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC3111046/]
• 22. The advantage of using SNP array in clinical testing ... [https://www.sciencedirect.com/science/article/abs/pii/S2210776213001245]
• 23. What are the Pros and Cons of FISH, aCGH and NGS? [https://www.enzo.com/note/what-are-the-pros-and-cons-of-fish-acgh-and-ngs/]
• 24. Which Cytogenetic Technique Is Right for You? [https://www.thermofisher.com/blog/behindthebench/which-cytogenetic-technique-is-right-for-you/]
• 25. Comprehensive performance comparison of high-resolution ... [https://bmcgenomics.biomedcentral.com/articles/10.1186/s12864-017-3658-x]
• 26. Array-based comparative genomic hybridization analysis of ... [https://www.nature.com/articles/gim200840]
• 27. Evaluation of the Effectiveness and Accuracy of Non- ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12383400/]
• 28. Non-invasive preimplantation genetic testing: a literature ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC9355447/]
• 29. Effectiveness of non-invasive chromosomal screening for ... [https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2023.1036467/full]
• 30. Improving embryo ploidy prediction: a machine learning ... [https://www.sciencedirect.com/science/article/abs/pii/S147264832500611X]
• 31. A foundational model for in vitro fertilization trained on 18 ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12254344/]
• 32. Non-invasive selection for euploid embryos: prospects and ... [https://www.sciencedirect.com/science/article/pii/S1472648325002846]
• 33. MAIA platform for routine clinical testing: an artificial ... [https://www.nature.com/articles/s41598-025-17755-y]
• 34. Predicting pregnancy outcomes in IVF cycles: a systematic ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12476623/]
• 35. Non-invasively predicting euploidy in human blastocysts via ... [https://rbej.biomedcentral.com/articles/10.1186/s12958-024-01302-x]
• 36. Deep learning applications for human embryo assessment ... [https://www.frontiersin.org/journals/reproductive-health/articles/10.3389/frph.2025.1549642/full]
• 37. iDAScore - Intelligent data analysis for embryo evaluation [https://www.vitrolife.com/our-products/idascore-intelligent-data-analysis-for-embryo-evaluation/]
• 38. Predictive Ability of an Objective and Time-Saving Blastocyst ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12292090/]
• 39. Improving Deep Learning-Based Algorithm for Ploidy ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11327420/]
• 40. Research progress of time-lapse imaging technology and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10519478/]
• 41. Time-Lapse Imaging in IVF: Bridging the Gap Between ... [https://www.mdpi.com/1422-0067/26/19/9609]
• 42. Morphokinetic analysis of in vitro fertilization outcomes [https://www.sciencedirect.com/science/article/pii/S0015028223006180]
• 43. Merging synthetic and real embryo data for advanced AI ... [https://www.nature.com/articles/s41598-025-94680-0]
• 44. A comparison of 12 machine learning models developed to ... [https://research.birmingham.ac.uk/en/publications/a-comparison-of-12-machine-learning-models-developed-to-predict-p]
• 45. Accurate machine learning model for human embryo ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12640390/]
• 46. The impact of euploid blastocyst morphology and maternal ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8995226/]
• 47. Effect of age and morphology on sustained implantation ... [https://www.sciencedirect.com/science/article/pii/S1472648321002923]
• 48. End-to-end deep learning for recognition of ploidy status ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8324635/]
• 49. Embryo ploidy status classification through computer ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10461251/]
• 50. Non-invasive prediction of human embryonic ploidy using ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11541425/]
• 51. The Effect of Advanced Maternal Age on Embryo ... [https://www.frontiersin.org/journals/endocrinology/articles/10.3389/fendo.2019.00686/full]
• 52. Accurate identification of abnormal ploidy using an artificial ... [https://pubmed.ncbi.nlm.nih.gov/40989216/?utm_source=FeedFetcher&utm_medium=rss&utm_campaign=None&utm_content=10MSL6XygFtCE0XpaQpog-pEt8342fXkY9lzrR7q6H0wVRyEa2&fc=None&ff=20250924085231&v=2.18.0.post9+e462414]
• 53. Human embryos harbor complex mosaicism with broad ... [https://www.nature.com/articles/s41421-024-00719-3]
• 54. Spectrum of Various Mosaicism Types According to Female Age [https://pmc.ncbi.nlm.nih.gov/articles/PMC12562040/]
• 55. Comprehensive benchmarking and guidelines of mosaic ... [https://www.nature.com/articles/s41592-023-02043-2]
• 56. The nature of embryonic mosaicism across female age ... [https://www.sciencedirect.com/science/article/pii/S2666334123000399]
• 57. FEMI helps predict IVF outcomes with greater accuracy [https://www.news-medical.net/news/20250715/FEMI-helps-predict-IVF-outcomes-with-greater-accuracy.aspx]
• 58. Stability and reliability of artificial intelligence models in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12494150/]
• 59. Accurate machine learning model for human embryo ... [https://link.springer.com/article/10.1007/s10815-025-03585-4]
• 60. Deep learning applications for human embryo assessment ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12011738/]
• 61. Comparing performance between clinics of an embryo ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC10440335/]
• 62. Non-invasive prediction of human embryonic ploidy using ... [https://www.sciencedirect.com/science/article/pii/S2589537024004760]
• 63. Intelligent assisted reproduction: Innovative applications of ... [https://www.sciencedirect.com/science/article/pii/S3050474025000242]
• 64. Machine learning for prediction of euploidy in human ... [https://www.sciencedirect.com/science/article/pii/S001502822102238X]
• 65. Application of a methodological framework for the ... - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11783712/]
• 66. Optimizing Non‐Invasive PGT‐A: A Multi‐Factorial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12529454/]
• 67. Automatic ploidy prediction and quality assessment of human ... [https://pubmed.ncbi.nlm.nih.gov/39237547/]
• 68. explainable AI for embryo ploidy prediction and patient- ... [https://pubmed.ncbi.nlm.nih.gov/38963605/]
• 69. A comparison of 12 machine learning models developed to ... [https://pubmed.ncbi.nlm.nih.gov/36825452/]
• 70. Comparing performance between clinics of an embryo ... [https://link.springer.com/article/10.1007/s10815-023-02871-3]
• 71. Abstract (COGI 2024): Artificial intelligence (AI) models ... [https://futurefertility.com/en/resources/abstract-cogi-2024-artificial-intelligence-ai-models-predict-blastocyst-development-and-ploidy-status-from-images-of-mature-metaphase-ii-mii-oocytes/]
• 72. Deep learning classification integrating embryo images with ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12095659/]
• 73. Predictive models for live birth outcomes following fresh ... [https://translational-medicine.biomedcentral.com/articles/10.1186/s12967-025-07045-6]
• 74. Deep-learning model for embryo selection using time- ... [https://www.nature.com/articles/s41598-025-10531-y]
• 75. WO2025076918A1 - 一种非侵入式整倍体预测方法及系统 [https://patents.google.com/patent/WO2025076918A1/zh]
• 76. AI在预测胚胎健康方面展现出前景，无需侵入性检测- AI与医疗健康 ... [https://www.familydoctor.cn/news/ai-yuce-peitai-jiankang-qianjing-wuxu-qinru-xing-jiance-11568.html]
• 77. AI算法非侵入性筛查试管婴儿胚胎 [http://health.people.com.cn/n1/2022/1226/c14739-32593521.html]
• 78. 延时AI 模型增强IVF 胚胎选择 [https://developer.nvidia.cn/blog/time-lapse-ai-model-enhances-ivf-embryo-selection/]
• 79. AI預測胚胎染色體整倍性！提高試管嬰兒植入成功率 [https://www.e-stork.com.tw/article/view/20533]
• 80. Comparative analysis of cumulative live birth rates in ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12224449/]
• 81. Association between a morphokinetic ploidy prediction ... [https://www.sciencedirect.com/science/article/pii/S0015028223005940]
• 82. Association between a morphokinetic ploidy prediction ... [https://research.birmingham.ac.uk/en/publications/association-between-a-morphokinetic-ploidy-prediction-model-risk-/]
• 83. Segmental aneuploidy in human blastocysts: a qualitative and ... [https://rbej.biomedcentral.com/articles/10.1186/s12958-019-0515-6]
• 84. Accurate identification of abnormal ploidy using an artificial ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12453672/]
• 85. Assessing the Viability of Segmental Aneuploid Embryos [https://pmc.ncbi.nlm.nih.gov/articles/PMC12155557/]
• 86. In the Pathway to Non-Invasive Techniques [https://www.mdpi.com/2073-4425/15/12/1613]