Enhancing Clinical IVF Embryo Selection through the Integration of Artificial Intelligence and Bayesian Statistics

Abstract

Despite major advancements in IVF technologies, the success rate of IVF remains low at around 35% and the main challenge of clinical IVF is embryo selection. Our approach to improve embryo selection leverages the advantages of AI, its ability to analyzing large and complex datasets, and Bayesian statistics, its ability to robustly infer causal relationships in complex problems. First, we enhanced the precision and efficiency of Time-Lapse Microscopy (TLM) image analysis through the implementation of Computer Vision (CV) techniques. We developed AI algorithms to automate the extraction of morphokinetic features from TLM movies of human embryos, providing a more objective evaluation of embryo quality. Second, we integrated these CV networks into BlastAssist, a pipeline designed to measure comprehensive, quantitative, and clinically-relevant features in IVF. To validate our pipeline, we conducted detailed comparisons between BlastAssist measurements and manual assessments by human experts, annotations from embryologists during routine treatments, outcomes of single embryo transfer (SET) cycles, and live birth outcomes of transferred embryos. Lastly, using the unprecedentedly large dataset generated by the BlastAssist pipeline in conjunction with electronic health record (EHR) data from three IVF labs, we constructed probabilistic graphical models (PGMs) for the complex IVF process across three key stages: ovarian stimulation, fertilization, and embryo development. These graphical models elucidate causal relationships among variables in clinical IVF, providing valuable insights for directing future clinical research.Overall, this project has the potential for significant clinical impact. The BlastAssist pipeline has the potential as a powerful tool to streamline the IVF image analysis process and assist embryologists in embryo selection. The integration of AI and Bayesian statistics can enhance our understanding of this complex process and help us identify key clinical predictors and improve clinical outcomes.