October 24th, 2022
Objective: We previously developed a machine learning model for selecting an individualized gonadotropin dose for ovarian stimulation based on patient similarity matching . The purpose of this study was to evaluate this methodology on a large and diverse dataset to retrospectively determine associations between machine learning predictions, patient outcomes, and total gonadotropin used.
Materials and Methods: We analyzed 365,473 autologous retrieval cycles from 2014-2019 in the Society for Assisted Reproductive Technology Clinical Outcomes Reporting System (SART CORS). Cycle information included patient age, BMI, AMH, total FSH dose, cycle length, and outcomes. The cumulative live birth rate (CLBR) for each retrieval cycle was calculated as at least one live birth from all linked embryo transfers. For cycles with blastocyst transfers (n = 161,035), the total blastocysts was defined as the number of embryos transferred plus embryos frozen. A K-nearest neighbors model was trained on all cycles to identify the 500 most similar cycles to a patient-of-interest using age, BMI, and AMH. For each cycle in the dataset, a patient specific dose-response curve was created by fitting a constrained second order polynomial to 2PNs retrieved relative to the average daily dose of FSH (the total FSH divided by the number of cycle days) across that patient’s 500 neighbors. Each patient’s individual dose-response curve was used to determine if there was an optimal daily dose that maximized the predicted 2PNs (called dose-responsive) or if the dose-response curve showed that the predicted 2PNs were approximately constant across the dose range (called flat-responsive).
Results: 27% of the cycles were identified as dose-responsive, while 73% were identified as flat-responsive. Dose-responsive patients who received an optimal daily dose had, on average, 1.2 more 2PNs, 0.5 more blastocysts, and 4% higher CLBR using 840 IU’s less of total FSH compared to propensity-matched patients with non-optimal doses. Flat-responsive patients who received a low daily dose (below the median of their neighbors) had, on average, 0.3 more 2PNs, 0.3 more blastocysts, and 3% higher CLBR using 2,020 IU’s less of total FSH compared to propensity-matched patients with a high daily dose.
Conclusions: Using the SART CORS dataset, this study shows that patient similarity modeling for selecting gonadotropin doses may be associated with improved ovarian stimulation outcomes and a reduction in total gonadotropin used. While propensity matching ensured covariates (age, BMI, and AMH) were similarly distributed when comparing outcomes between groups, future prospective studies are needed to determine potential clinical benefit of this machine learning approach.
Impact Statement: Machine learning tools to aid in decision making during ovarian stimulation show potential for improving outcomes including live birth while reducing the total amount of FSH.
References:  Loewke KE, Nutting VI, Cho JH, Hoffman DI, Weckstein LN, Levy M. An interpretable machine learning model for individualized protocol selection and gonadotropin dose selection during ovarian stimulation. Fertil Steril. 2021 Sep 1;116(3):e174.