An interpretable and generalizable machine learning model for optimizing day of trigger during ovarian stimulation

Background: The goal of ovarian stimulation during in vitro fertilization (IVF) is to promote multifollicular development in order to retrieve multiple high-quality eggs. One of the most important decisions during this process is the timing of the trigger injection. Triggering too early may prevent eggs from reaching maturity, while triggering too late may result in post-mature eggs and risk of hyperstimulation. Machine learning models can support in making this decision, but should be interpretable so the basis for their recommendations are understood, and generalizable to perform well on data from different clinics.

Objective: To develop an interpretable and generalizable machine learning model for determining the optimal day of trigger during ovarian stimulation.

Materials and Methods: Historical, de-identified electronic medical record (EMR) data were collected for IVF retrieval cycles started between 2014-2020 from 3 different IVF clinics in the United States, for a total of approximately 18,000 cycles. EMR records contained cycle outcomes, daily measurements of estradiol (E2), and daily follicle counts and sizes. Data were split into train and test datasets stratified by site. Models were developed to predict the number of MII eggs that will be retrieved if triggering a patient in the current day (today) compared to the next day (tomorrow), with a goal of having model interpretability. To predict MII eggs if triggering today, a linear regression model was developed using follicle counts and E2 levels measured on the day of trigger. Next, a follicle forecasting model was developed to forecast follicle counts and E2 levels one day in the future. These forecasted values were then used as inputs into the linear regression model to predict MII eggs if triggering tomorrow. Together, these models were used to recommend an optimal day of trigger. By comparing recommendations to actual trigger days, patients were classified as having an early, on-time, or late trigger. Propensity score matching was used to match patients in the early/late group with patients in the on-time group.

Results: On the test dataset, the linear regression model predicted MII eggs on the day of trigger with a mean absolute error (MAE) of 2.78 eggs and an R² of 0.65. Combining the linear regression and follicle forecasting models, the model predicted next-day MII eggs with an MAE of 3.06 and an R² of 0.63. Possible early and late triggers were identified in 56% and 5% of cycles, respectively. Patients with early/late triggers had on average 2.6 fewer MIIs, 1.9 fewer 2PNs, and 1.0 fewer usable blastocysts when compared to propensity-matched patients with on-time triggers.

Conclusions: We developed an interpretable machine learning approach for optimizing the day of trigger. Clinic-specific results showed comparable levels of accuracy indicating generalizability over a diverse set of patients. Our results indicate that our model can be used as a clinical decision support tool to potentially optimize clinical and laboratory outcomes for a significant number of patients, as well as help with patient counseling.

Financial Support: This study was supported by Alife Health.