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

Most artificial intelligence (AI) systems for embryo selection in in-vitro fertilization (IVF) rely on single-instance learning (SIL), evaluating each embryo independently from morphological features to predict implantation or live birth. However, this approach does not reflect how embryologists operate in practice by comparing embryos within a cohort to select the most viable one for transfer. Recent evaluations of commercial SIL models have revealed high variability in embryo rankings, limited agreement with expert assessments, and inconsistency across datasets, raising concerns about their clinical reliability and trustworthiness. To address these limitations, we investigated whether a multi-instance learning model, Cohort-AI, could improve the stability, reliability, and clinical agreement of AI-assisted embryo selection by evaluating embryos in the context of their full cohort. Using a retrospective design, we trained and validated 50 replicate Cohort-AI models and 50 SIL models on embryo image datasets from two fertility centers (Massachusetts General Hospital and Weill Cornell Medicine), comprising over 11,000 embryos. We compared ranking consistency, critical error rates, concordance with clinical embryo transfers, and live birth outcomes across both approaches. Focusing on shared ranking endpoints rather than predictive accuracy allowed comparison of SIL and Cohort-AI despite their different training objectives (individual embryo versus cumulative cycle outcome). Cohort-AI showed higher ranking stability than SIL models (Kendall's W ∼0.93 vs. ∼0.35), reduced critical error rates (∼2% vs. ∼15%), and showed closer agreement with clinician-selected embryos (transfer rate: 68.7% vs. 44.7%). Embryos selected by Cohort-AI also correlated with higher live birth rates and maintained consistent performance across institutions, in this two-center retrospective evaluation, without additional adaptation to the external site. In this retrospective, two-center evaluation, the cohort-aware approach improved ranking stability and reduced conservative critical-error rates relative to SIL. Prospective, multi-center studies are warranted to establish real-world clinical utility.