The research, led by Prudhvi Thirumalaraju, Manoj Kumar Kanakasabapathy, and Hadi Shafiee of Brigham and Women’s Hospital and Harvard Medical School, together with collaborators at Weill Cornell Medicine, introduces a model called Cohort-AI. Rather than evaluating embryos one by one, Cohort-AI uses a machine learning framework known as multi-instance learning to assess an entire patient’s set of embryos simultaneously, predicting the cumulative live-birth outcome of the cohort and ranking each embryo by its relative contribution to that outcome. The approach mirrors the way embryologists reason: an embryo that looks average on its own might be the best option in a weak cohort, while a good-looking embryo might rank lower in a set full of excellent candidates.

The motivation for the work comes from an uncomfortable finding in recent literature. A previous evaluation of eight commercial AI algorithms for embryo ranking found that agreement between the algorithms and human embryologists was lower than expected, and in some cases the models disagreed with each other. Most strikingly, two of the commercial models produced rankings that were statistically indistinguishable from random chance. Even when the models were fine-tuned on center-specific data, the discrepancies persisted. The problem, the researchers argue, is a phenomenon known in machine learning as underspecification: when many different internal solutions can achieve similar overall accuracy, small variations in training—such as the random seed used to initialize the network’s weights—can produce large, unpredictable changes in individual decisions, even while headline performance metrics look stable.

To test whether a cohort-aware design could tame this instability, the team trained fifty replicate versions of both a conventional single-instance convolutional neural network and the new Cohort-AI model, differing only in their random initialization. All training used more than 10,000 Day-5 embryo images from 1,258 patients at Massachusetts General Hospital, captured with time-lapse incubators. A separate dataset of 648 embryos from 53 patients at Weill Cornell was held out entirely as an independent external test set, with no tuning or adaptation of any kind. The researchers then measured how consistently the fifty replicate models of each type ranked the same embryos for the same patients.

The results were stark. On the Massachusetts General Hospital test set of 92 patient cohorts, the replicate Cohort-AI models achieved an average Kendall’s W—a statistical measure of rank-order agreement ranging from zero, meaning no agreement, to one, meaning perfect agreement—of approximately 0.94. The single-instance models managed only about 0.36. On the external Cornell data, the pattern held: roughly 0.91 for Cohort-AI versus 0.34 for the conventional approach, with both differences highly statistically significant. In practical terms, the conventional models, trained identically except for a random starting point, frequently disagreed with themselves about which embryo was best, while the cohort-aware models produced nearly identical rankings across all fifty replicates.

Consistency translated directly into fewer dangerous mistakes. The team defined a critical error as a case in which a model ranked a degenerate, grade 1 embryo as the top choice even though a viable blastocyst of grade 3 or better was present in the same cohort. On the internal test set, single-instance models made such errors at an average rate of about 12.4 percent, with individual models ranging from roughly 4 to 22 percent. Cohort-AI cut that rate to about 2.3 percent. On the external Cornell data the gap widened further: about 17.3 percent for the conventional models versus 1.24 percent for Cohort-AI. Notably, the variance in error rates across replicate models also collapsed for the cohort-aware approach, and formal testing showed that this variance did not significantly increase when the model was applied to the unfamiliar Cornell data—evidence that the stability survived a genuine distribution shift between two different clinics and microscope models.

The study also probed what the model had actually learned. Using attention mechanisms—the component of the network that assigns weights to different inputs when forming a decision—the researchers found that Cohort-AI consistently concentrated its attention on embryos of higher morphological grade, as judged by the modified Gardner grading system used in clinics, and assigned the lowest attention to empty wells in the culture dish. Feature-space visualizations showed that replicate Cohort-AI models clustered tightly together in how they represented embryos from the same patient, whereas the conventional models diverged widely. The authors are careful to note that attention weights are treated as a descriptive weighting mechanism, not as causal proof of embryo importance, but the alignment with embryologist grading suggests the model’s comparative reasoning resembles human judgment rather than exploiting spurious shortcuts.

Perhaps most importantly for clinical credibility, the stability did not come at the cost of performance. In retrospective comparisons against historical clinical outcomes, Cohort-AI’s top-ranked embryos matched the embryos clinicians actually transferred 68.7 percent of the time at Massachusetts General Hospital, compared with 44.7 percent for the single-instance models, and the live-birth rate associated with those top selections was 44.8 percent versus 41.0 percent—both above the center’s baseline of 35.1 percent for the evaluated patients. At Cornell, Cohort-AI again showed higher transfer and live-birth rates with at least twice the consistency. When the analysis was restricted to cycles known to contain at least one successful live birth, the median Cohort-AI model produced 29 live births from about 35 transfers on the internal data, while the median conventional model produced only 16 from about 23 transfers. The lowest-performing Cohort-AI replicates performed comparably to the highest-performing conventional models, meaning clinicians would no longer need to gamble on which replicate they happened to deploy.

A single-patient case study crystallized why these differences matter. For one patient with three high-quality blastocysts, one moderate blastocyst, and two degenerate embryos, consensus among five embryologists was unambiguous. Yet 33 of the 50 conventional models ranked a degenerate embryo above a high-quality one, 31 ranked a high-quality embryo last, and seven—including the three models with the highest validation accuracy—placed a degenerate embryo first. Pairwise rank correlations among the conventional models averaged just 0.27, with some pairs producing exactly opposite orderings. All fifty Cohort-AI models, by contrast, ranked the high-quality embryos at the top and the degenerate ones at the bottom, with an average pairwise correlation of 0.99. The two embryos Cohort-AI consistently ranked highest were among those that had resulted in live births. The lesson, the authors argue, is that strong validation accuracy is an insufficient proxy for clinical reliability in ranking tasks.

The implications extend beyond embryology. The authors contend that embryo selection is inherently a comparative problem, not a binary classification, and that validation frameworks designed for diagnostic AI—where each sample has a clear ground truth—are poorly suited to it. They call for standardized benchmarks that measure not only accuracy but also ranking reproducibility, cross-site robustness, and behavior across software updates, noting that AI adoption in IVF has surged from roughly a quarter of clinics in 2022 to more than half in 2025 even as many clinicians report low confidence in interpreting these systems. The team is candid about limitations: the study is retrospective, the associations with live birth are correlational rather than causal, both centers used the same family of time-lapse incubators, and no blinded comparison against embryologist rankings was performed. Prospective, multi-center trials remain the necessary next step. But the central message is clear and potentially field-changing: for AI to become dependable infrastructure in IVF, rankings must remain predictable as data and software evolve—and building the comparison into the model itself, rather than leaving it out, appears to be a powerful way to get there.

Subject of Research: A multi-instance learning AI model for stable and reliable AI-assisted embryo selection in IVF

Article Title: Cohort-AI: A multi-instance learning approach for improved stability and reliability in AI-assisted embryo selection

Article References: Thirumalaraju, P., Kanakasabapathy, M. K., Kandula, H., Kandula, T., Katkuri, A. V. R., Cipriano, C., Malmsten, J. E., Zaninovic, N., Bormann, C. L., & Shafiee, H. (2026). Cohort-AI: A multi-instance learning approach for improved stability and reliability in AI-assisted embryo selection. Heliyon, 12(15), Article e45457. https://doi.org/10.1016/j.heliyon.2026.e45457

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45457

Keywords: in vitro fertilization, embryo selection, artificial intelligence, multi-instance learning, deep learning, reproductive medicine, Kendall's W, model stability, underspecification, attention mechanisms, clinical decision support, time-lapse imaging

The study, published in Reproductive Sciences by Zhenya Yuan, Jun Shao, Ya Wen and colleagues working across reproductive medicine centers in Xuzhou, Nantong and Kunming, addresses a problem that is particularly acute in China, where PGT-A is subject to strict regulatory limitations. Without access to genetic testing, embryologists must choose which embryos to transfer based on appearance and development alone, yet visual assessment is a notoriously imperfect proxy for chromosomal health. An embryo can look flawless under the microscope and still carry an extra or missing chromosome, while a seemingly mediocre embryo may be genetically normal and fully capable of producing a live birth.

The motivation for avoiding biopsy is not merely regulatory. Although PGT-A is widely used, concerns persist that removing cells from a blastocyst, typically from the trophectoderm that later forms the placenta, could compromise the embryo’s developmental potential. The new study’s authors note that this possibility of harm, combined with the regulatory environment, made a non-invasive predictive approach a clinical necessity rather than a convenience. Their solution draws on data that fertility clinics already collect routinely, meaning the model could in principle be deployed without any change to laboratory workflow.

Technically, the researchers constructed their model using logistic regression, a statistical machine learning method that estimates the probability of a binary outcome, in this case whether a blastocyst is euploid or aneuploid. The team describes the construction as a two-step logistic regression analysis, in which predictor variables were combined to generate a risk score for each embryo. The inputs to the model fell into four categories: the results of PGT-A performed on the training embryos, which served as the ground truth labels; the dynamic development characteristics of the embryos, capturing how quickly and smoothly they progressed through early cleavage stages; the morphology of the blastocysts themselves, including the conventional features embryologists assess under the microscope; and the age of the female patient, which is one of the strongest known determinants of aneuploidy risk.

This combination of features is significant because each captures a different dimension of embryo quality. Morphokinetic data, the timing of developmental events, reflects the pace and coordination of cell divisions, and previous research has linked deviations from normal timing to chromosomal abnormalities. Blastocyst morphology, graded by systems such as the Gardner classification, summarizes the expansion of the embryo and the quality of the inner cell mass and trophectoderm. Female age contributes the statistical prior that chromosomal error rates rise steeply as oocytes age. By weaving these signals together mathematically, the model effectively formalizes what skilled embryologists do intuitively, but does so consistently and with quantified confidence.

The performance figures reported for the model are striking for a non-invasive approach. The area under the receiver operating characteristic curve, a standard measure of a classifier’s ability to distinguish between positive and negative cases across all decision thresholds, reached 0.877. When the model was applied to an independent data set, meaning embryos that played no role in training it, it achieved an accuracy rate of 77.35 percent, a precision rate of 77.49 percent, and a recall rate of 72.55 percent. The f-score, which balances precision and recall into a single number, was 74.94 percent. The model correctly identified aneuploid embryos, the true negative rate from the euploid perspective, 81.55 percent of the time, while producing a false positive rate of 18.45 percent and a false negative rate of 27.45 percent.

Perhaps the most clinically meaningful comparison in the study pits the machine learning model against the tools embryologists actually use. The model’s false positive rate in predicting euploid blastocysts was significantly lower than that of the Gardner grading system, the standard morphology-based classification, and also lower than KIDScore D5, a commercially available time-lapse algorithm that scores embryos based on their developmental dynamics. In practical terms, this means the model was less likely than conventional methods to flag an embryo as euploid when it was not, reducing the risk that patients undergo transfers destined to fail or miscarry.

The researchers went beyond retrospective prediction and examined what happened when the model was used to make real selection decisions. When blastocysts for frozen embryo transfer were chosen using the model’s predictions, the transfer outcomes improved significantly compared with cycles in which embryos were selected by Gardner grade alone. This is a crucial distinction, because a model can score well on statistical metrics yet fail to change clinical results. Here, the improvement in frozen embryo transfer outcomes suggests that the model’s predictions translate into genuinely better choices about which embryos to place into patients.

The study builds on a growing body of work applying artificial intelligence to embryo assessment. Earlier efforts by some of the same authors developed an artificial intelligence model for predicting blastocyst euploidy specifically within PGT-A treatment cycles, where genetic results are available to train and validate algorithms. Other groups have compared deep learning-based embryo selection against manual morphology assessment in randomized trials, and meta-analyses of thousands of embryos have examined how morphological and morphokinetic features correlate with aneuploidy. Time-lapse imaging studies have identified specific dynamic events, such as blastocyst contractions, that associate with chromosomal abnormality and reduced implantation potential. The new work extends this lineage into territory where genetic ground truth is unavailable at the point of care, which is precisely where the clinical need is greatest.

The implications reach well beyond China’s regulatory landscape. Even in countries where PGT-A is freely available, the procedure adds cost, requires embryo biopsy, and remains debated, with large randomized trials yielding mixed conclusions about whether it improves live birth rates per transfer cycle started. A validated non-invasive predictor could serve as a triage tool everywhere, helping clinics prioritize embryos for transfer or for biopsy, sparing patients unnecessary procedures, and potentially reducing the number of embryos subjected to manipulation. Because the model relies on development dynamics, morphology and maternal age, data captured in virtually every modern IVF laboratory, adoption barriers would be low compared with approaches requiring new imaging hardware or genetic assays.

Caution is still warranted. A model with 77 percent accuracy and a 27 percent false negative rate is a powerful aid to decision-making, not a replacement for genetic diagnosis when a definitive answer is required, and the authors themselves frame the tool as a way to select euploid embryos as far as possible rather than with certainty. External validation across diverse patient populations, laboratory protocols and imaging systems will be essential before widespread deployment, and the field has seen predictive models falter when moved between centers. Nevertheless, the study represents a concrete step toward a long-sought goal: choosing the best embryo for transfer using only what the embryo reveals about itself, non-invasively, as it grows. For the millions of couples undergoing IVF worldwide, an algorithm that quietly reads the developmental story of each blastocyst and tips the odds toward a healthy pregnancy is a development worth watching closely.

Subject of Research: Machine learning prediction of euploid blastocysts in IVF without preimplantation genetic testing for aneuploidy

Article Title: The Building and Application of a Machine Learning Model on Predicting Euploid Blastocysts in IVF Treatments Without PGT-A

Article References: Yuan, Z., Shao, J., Wen, Y., Yuan, M., Bai, M., Song, X., & Huang, X. (2026). The Building and Application of a Machine Learning Model on Predicting Euploid Blastocysts in IVF Treatments Without PGT-A. Reproductive Sciences. https://doi.org/10.1007/s43032-026-02230-4

Image Credits: AI Generated

DOI: 10.1007/s43032-026-02230-4

Keywords: IVF, machine learning, euploid blastocyst, PGT-A, aneuploidy, embryo selection, logistic regression, morphokinetics, blastocyst morphology, frozen embryo transfer, reproductive medicine, non-invasive prediction

To understand why this matters, it helps to look at how CAPA-IVM differs from conventional in vitro fertilization. Standard IVF relies on controlled ovarian stimulation, in which women receive daily injections of gonadotropins to mature many eggs inside the body before retrieval. That stimulation carries a well-known danger: ovarian hyperstimulation syndrome, or OHSS, a potentially serious condition driven by vascular endothelial growth factor and excessive luteal stimulation. Strategies such as triggering final egg maturation with a gonadotropin-releasing hormone agonist instead of hCG, and freezing all embryos for transfer in a later cycle, have reduced but never fully eliminated this risk. IVM takes a more radical approach by skipping stimulation almost entirely. Immature eggs are retrieved from small follicles of just two to eight millimeters and matured in the laboratory under low-dose hormones, followed by intracytoplasmic sperm injection.

Conventional IVM, however, has historically lagged behind IVF in success rates, partly because the retrieved cohort of eggs matures in a dys-synchronized fashion. CAPA-IVM addresses this with a pre-maturation phase lasting roughly 24 hours, in which C-type natriuretic peptide keeps the eggs in meiotic arrest by maintaining elevated cyclic AMP levels inside the surrounding cumulus cells. This pause allows nuclear and cytoplasmic maturation to proceed in step. A subsequent 30-hour maturation phase with follicle-stimulating hormone and amphiregulin then drives development forward. A landmark randomized controlled trial published in 2020 compared CAPA-IVM with conventional IVF in 546 women with high antral follicle counts. Although live birth rates after the first transfer were somewhat lower with CAPA-IVM, at 35.2 percent versus 43.2 percent, the technique produced zero cases of OHSS, required far less medication, and showed no increase in neonatal complications or developmental delays in follow-up studies of children born through the method.

Despite these clinical advantages, CAPA-IVM faces a practical bottleneck: embryo selection. Because the technique yields fewer embryos than stimulated IVF, choosing the single best candidate for transfer carries extra weight. Since the late 1990s, time-lapse monitoring has offered a way to watch embryos develop continuously inside the incubator, photographing them at regular intervals without removing them from their culture environment. This produces a rich record of morphokinetics, the exact timing of events such as the first cell divisions, compaction, and the onset of blastulation. Around the world, artificial intelligence has been layered onto this data stream, with deep learning systems such as IVY and iDAscore demonstrating that algorithms can rank embryos and predict implantation with accuracy that rivals or exceeds human embryologists.

What has been missing, until now, is any AI model built specifically for CAPA-IVM embryos, whose developmental trajectories may differ from those of stimulated IVF embryos because they originate from small, unstimulated follicles. The Vietnamese team, a collaboration between the Hope Research Center at My Duc Hospital and the Laboratory of AI and Computer Vision at the University of Arkansas, set out to fill that gap. Their retrospective cohort study ran from March 2021 to December 2023 and included 81 women aged 18 to 37 with polycystic ovarian morphology who underwent CAPA-IVM. Embryos were cultured to the blastocyst stage in an Astec time-lapse incubator capturing images every 15 minutes, then cryopreserved for later frozen transfer.

The researchers assembled a carefully balanced dataset of 100 embryos: 50 that resulted in a live birth at or beyond 24 weeks of gestation with a healthy newborn, and 50 that did not, matched by comparable female age. From each embryo’s time-lapse video, they extracted frames using Python and OpenCV and annotated both morphological grades and a battery of morphokinetic timings, following ESHRE recommendations and the Istanbul Consensus guidelines. The predictor variables included the times to pronuclear appearance and fading, the two-cell, four-cell, and five-cell stages, the start of compaction, the morula stage, and the start of blastulation, alongside embryo grade, the proportion of mature metaphase II oocytes, maternal age and body mass index, infertility type, and the total number of embryos frozen per cycle.

The core experiment tested a deceptively simple question: does an algorithm need all eight morphokinetic parameters to predict live birth, or can a stripped-down set of four do the job just as well? Four supervised learning algorithms were pitted against each other under both feature specifications: logistic regression, decision tree, gradient boosting, and XGBoost. Hyperparameters were tuned through exhaustive grid search with repeated stratified five-fold cross-validation, repeated ten times, and the area under the receiver operating characteristic curve served as the primary metric. The results revealed an instructive pattern. With the reduced four-parameter feature set, the decision tree model led the field, reaching 80 percent test accuracy and a ROC AUC of 0.775, with perfect recall on live birth cases. With the full eight-parameter set, XGBoost took the crown, matching the 80 percent accuracy but pushing the ROC AUC to 0.82.

The interpretability analysis proved as revealing as the raw performance numbers. Using SHAP values, a technique that quantifies each feature’s contribution to individual predictions, the researchers found that embryo morphological grade dominated both models, with Grade 3 embryos far less likely to end in live birth. In the reduced model, the timing of the four-cell stage emerged as a key early marker of developmental competence, along with the MII oocyte rate and body mass index. In the full model, later events such as the start of compaction, the two-cell timing, and the five-cell timing added predictive value, suggesting that algorithms capable of handling nonlinear interactions can squeeze extra signal from dense developmental trajectories. Notably, the two best models delivered performance competitive with published IVF embryo-ranking systems, which typically report ROC AUC values between 0.65 and 0.75, a striking result for a first attempt in the IVM setting.

The practical implications reach well beyond the statistics. If four carefully chosen timing markers capture most of the clinically relevant information, embryology laboratories could dramatically reduce the annotation workload that comprehensive morphokinetic tracking demands, improving reproducibility across centers and lowering the barrier to adopting AI-assisted selection. Simpler models with fewer, more interpretable inputs are also easier for clinicians to trust and integrate into daily decision-making, particularly for patients with polycystic ovary syndrome, who stand to benefit most from CAPA-IVM’s hormone-light approach. The authors caution, however, that this was a retrospective, single-center study with a modest sample of 100 embryos, some of which were nested within the same couples, and that external validation on larger, independent, multicenter datasets remains an essential next step before clinical deployment.

Even with those caveats, the study marks a meaningful convergence of two trends in reproductive medicine: the shift toward gentler, safer stimulation protocols, and the rise of machine learning as an objective arbiter of embryo quality. For the millions of women with PCOS or high antral follicle counts who face an elevated risk of ovarian hyperstimulation, a future in which an algorithm watches their embryos divide and flags the most promising candidate, all without a single day of heavy hormone injections, is moving from speculation toward the laboratory bench. As validation studies multiply, the humble decision tree and its gradient-boosted rival may become quiet workhorses of the modern fertility clinic, turning the silent choreography of cell division into a data-driven verdict on which embryo deserves its chance at life.

Subject of Research: Machine learning prediction of live birth outcomes for CAPA-IVM blastocyst embryos using time-lapse morphokinetic parameters

Article Title: Selection of capacitation in-vitro maturation (CAPA–IVM) blastocyst embryos by a machine learning prediction model using the morpho-kinetics

Article References: Dang, A., Nguyen, T., Lam, H., Nguyen, T., & Pham, H. (2026). Selection of capacitation in-vitro maturation (CAPA–IVM) blastocyst embryos by a machine learning prediction model using the morpho-kinetics. Results in Optics, Article 101165. https://doi.org/10.1016/j.rio.2026.101165

Image Credits: AI Generated

DOI: 10.1016/j.rio.2026.101165

Keywords: CAPA-IVM, in vitro maturation, embryo selection, machine learning, morphokinetics, time-lapse monitoring, XGBoost, decision tree, polycystic ovary syndrome, ovarian hyperstimulation syndrome, blastocyst, assisted reproduction