An online tool for predicting ovarian responses in unselected patients using dynamic inhibin B and basal antimüllerian hormone levels

Background Reliable predictive models for predicting excessive and poor ovarian response in controlled ovarian stimulation (COS) is currently lacking. The dynamic (Δ) inhibin B, which refers to increment of inhibin B responding to exogenous gonadotropin, has been indicated as a potential predictor of ovarian response. Objective To establish mathematical models to predict ovarian response at the early phase of COS using Δinhibin B and other biomarkers. Materials and methods Prospective cohort study in a tertiary teaching hospital, including 669 cycles underwent standard gonadotropin releasing hormone (GnRH) antagonist ovarian stimulation between April 2020 and September 2020. Early Δinhibin B was defined as an increment in inhibin B from menstrual day 2 to day 6 through to the day of COS. Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression with 5-fold cross-validation was applied to construct ovarian response prediction models. The area under the receiver operating characteristic curve (AUC), prevalence, sensitivity, and specificity were used for evaluating model performance. Results Early Δinhibin B and basal antimüllerian hormone (AMH) levels were the best measures in building models for predicting ovarian hypo- or hyper-responses, with AUCs and ranges of 0.948 (0.887–0.976) and 0.904 (0.836–0.945) in the validation set, respectively. The contribution of the early Δinhibin B was 67.7% in the poor response prediction model and 56.4% in the excessive response prediction model. The basal AMH level contributed 16.0% in the poor response prediction model and 25.0% in the excessive response prediction model. An online website-based tool (http://121.43.113.123:8001/) has been developed to make these complex algorithms available in clinical practice. Conclusion Early Δinhibin B might be a novel biomarker for predicting ovarian response in IVF cycles. Limiting the two prediction models to the high and the very-low risk groups would achieve satisfactory performances and clinical significance. These novel models might help in counseling patients on their estimated ovarian response and reduce iatrogenic poor or excessive ovarian responses.