Multimodal intelligent prediction model for in vitro fertilization.

Artificial intelligence has facilitated the automated selection of embryos and the prediction of pregnancy outcomes during in vitro fertilization (IVF), yet multimodal approaches remain underexplored-particularly for predicting multiple specific outcomes such as singleton pregnancy vs. multiple pregnancy, and miscarriage vs. live birth. In this study, we propose VaTEP (Video and Table model for Embryo Prediction), a multimodal embryo prediction model integrating time-lapse system (TLS) videos and tabular clinical variables. VaTEP is first pretrained on two pre-tasks (TLS video reconstruction and embryo developmental phase prediction) to fully capture the rich spatiotemporal dynamics and developmental information contained in the video, and further improved by a multiple frame sampling strategy and multitask learning framework. These designs enable VaTEP to estimate the probabilities of fetal heartbeat, singleton vs. multiple pregnancy, and miscarriage vs. live birth, promoting more informed embryo selection and outcome precognition. This helps reduce the risk of implantation failure by minimizing the chances of non-viable pregnancies, multiple gestations, and miscarriages. VaTEP offers a comprehensive and data-driven tool for personalized IVF decision-making, supporting safer and more effective reproductive treatment.