Using deep learning to predict the sex of human embryos.

Abstract

The existence of sex differences in human pre-implantation development remains an open question, with previous attempts based on human observations yielding inconclusive results. In this study, we combined manual annotation and deep learning analysis of a dataset comprising 515 time-lapse embryo movies to investigate whether birth sex influences early developmental dynamics. While manual assessment did not identify any developmental timing parameters that reliably distinguish male from female embryos, a deep learning model trained and tested on these videos achieves a statistically significant sex prediction accuracy of 61%. Importantly, our analyses identified the period after the eight-cell stage as critical for accurate prediction, indicating that subtle sex-related differences may begin to emerge around day 3 of human embryogenesis. Studying sex differences at this early stage may enhance our understanding of why some embryos fail to develop and why sex ratios can be skewed in the context of in vitro fertilization. More broadly, our findings open the possibility for an early, non-invasive detection tool that could assist in identifying and addressing sex-related embryonic developmental abnormalities.