Predicting the ovarian response: towards a determinant model and implications for practice.

Objective: To improve the reliability of prediction models for ovarian response to stimulation in ART.

Design: A multicenter retrospective cohort study.

Setting: Twelve reproductive centers.

Patients: A total of 25,854 controlled ovarian stimulations between 2005 and 2016, including cycles cancelled for inadequate response, were included.

Intervention(s): None.

Main Outcome Measure(s): Precision of the prediction of the number of oocytes at ovarian pickup and of cancellation rate for poor ovarian response.

Results: Both AMH and antral follicle count exhibit a non-linear effect on the oocyte yield, with a linear relationship after log-transformation. After adjustment for age, BMI, and center, ovarian response observed in a previous stimulation was found to be the best predictor, followed by AMH and AFC. The zero-inflated binomial negative model showed that predictors of cycle cancellation and number of oocytes at retrieval were different, and assimilating cancellation to zero oocyte greatly reduces the determination of the model. Our model was characterized by the best ever reached determination (R=0.505 for non-naïve women, 0.313 for all the women) and provided evidence of a very strong difference among centers. The results can be easily converted in the prediction of response levels (poor-medium-good-high). Finally, in case of partial report of the above predictors, we show that the univariate prediction based on the best predictor provides a good approximation.

Conclusion(s): A substantial improvement of the ovarian response prediction is possible in modelling the possible cancellation decision, followed by the oocyte retrieval itself, according to an appropriate model based on previous stimulation and non-linear effects of AMH and AFC.