and E2 levels, and evaluated in terms of R-squared (R2) and mean absolute error (MAE). Together, the two models enabled the prediction of the number of MII eggs when triggering on the last day of stimulation compared to one day in the future (i.e. trigger "today vs. tomorrow").RESULTS: The RNN predicted next-day follicles with MAE¼2.51 follicles and next-day estradiol levels with MAE¼375 IUs (16% of E2 level), for a test dataset comprising 20% of the data. The MAEs for next-day perbin follicle counts were 1. 96, 1.88, 1.36, 1.17, .89, and .60 follicles for bins <¼11mm, 12-14mm, 15-16mm, 17-18mm, 19-20mm, and >20mm, respectively. Multiple linear regression with recursive feature elimination identified baseline AFC, E2, and follicles <¼11mm, 12-14mm, 15-16mm, 17-18mm, and 19-20mm as significant predictors of MII eggs retrieved, and predicted MII eggs with R2¼0.62 and MAE¼3.11 eggs. When combining follicle and E2 forecasting with the linear model, the integrated approach predicted that 35% cycles may have resulted in more MII eggs if trigger had occurred one day later.CONCLUSIONS: We developed a machine learning approach for optimizing the day of trigger during ovarian stimulation. Our approach combined an RNN to forecast E2 and follicle counts with an interpretable linear model to predict the number of MII eggs retrieved. These models predicted that a significant number of cycles may have resulted in more MII eggs if trigger had occurred one day later. We acknowledge that for some cases, earlier trigger may have been chosen to prevent the risk of adverse clinical outcomes, which will be investigated in future work.IMPACT STATEMENT: We have developed a machine learning approach for forecasting E2 and follicle growth and predicting the number of MII eggs retrieved during ovarian stimulation, which may help with the decision of triggering "today vs. tomorrow".
