Trial emulation for validating the clinical efficacy of a foundational AI model in embryo selection

AI models for embryo selection often rely on correlational metrics that ignore clinical confounders. We present a target trial emulation framework to approximate causal effects for a foundational AI model, FEMI, for non-invasive embryo assessment using multi-center trial emulation (n = 4674). Propensity score matching established a robust association between FEMI-Ploidy and implantation failure showing an average treatment effect (ATE) of -0.131 (95% CI [-0.196, -0.066]) in the development cohort and -0.157 (95% CI [-0.254, -0.054]) in the external cohort. Comparative efficacy using S-Learner models demonstrated that a high-risk FEMI score carried a significantly stronger individual treatment effect (ITE) penalty on implantation compared to other scoring mechanisms (p < 0.0001). This superiority persisted after adjusting for maternal age, suggesting FEMI captures unique biological features. This causal framework establishes a rigorous standard for AI validation in IVF, providing the necessary pre-clinical justification for prospective randomized controlled trials.