The impact of clinical factors on the number of MII oocytes

Abstract Study question What is the impact of clinical factors on number of MII oocytes retrieved after ovarian stimulation? Summary answer Using selected clinical factors, Machine Learning models are capable of accurately predicting the number of MII oocytes. What is known already There are well known clinical factors that correlate with the number of MII oocytes that can be harvested as a result of ovarian stimulation. When deciding on an appropriate stimulation protocol and gonadotropin doses, gynecologist mainly take into consideration patient’s age and selected markers of ovarian reserve. The choice of protocol and dosage itself influences the outcome of the stimulation. What is still debated is to what extent each of the aforementioned factors influences the final outcome of the stimulation. Recent progress in Machine Learning now allows to develop predictive model that address this question. Study design, size, duration In the retrospective study, data collected between Nov 2014 and Jan 2021 was analyzed. Dataset consisted of 8428 ovarian stimulations undergone by 5692 patients at the Invicta Fertility Clinics. Eight clinical factors chosen by gynecologists were used in the modelling phase. Study’s inclusion criteria set boundaries on age (between 24 and 46 years), Anti-Müllerian hormone (AMH) (lower than 15 ng/ml) and the number of MII oocytes retrieved as a result of stimulation (fewer than 20). Participants/materials, setting, methods For MII oocytes prediction a Machine Learning model based on gradient boosting technique was created. The model contains 500 decision trees with 16 leaves and maximal depth of 64 nodes. Model’s performance was evaluated with median of absolute error and root mean square error (RMSE) metrics. To validate the model, a k-fold cross validation was used with 5 separate folds. Importance of aforementioned clinical factor in the model was assessed with Shapley values. Main results and the role of chance To predict the number of MII oocytes, following factors were included: AMH, antral follicle count on the start of stimulation (AFC), patient age, number of MII oocytes retrieved in the last stimulation (available for 37% of IVF cycles), daily gonadotropin dose in days 1-3 and 4-7, type of IVF protocol and the presence of Polycystic Ovary Syndrome (PCOS). Model achieved an RMSE of 3.38 and a median absolute error of 2.09. AMH was proved to have the greatest impact on the predicted number of MII, with an average impact on the model’s output magnitude (SHAP) equal to 1.40. AFC was identified as the second most important variable (SHAP=0.78). The number of MII oocytes from last stimulation (SHAP=0.21), daily gonadotropin dose used in days 1-3 of stimulation (SHAP=0.18), patient age (SHAP=0.17), daily gonadotropin dose days 4-7 of stimulation (SHAP=0.08), type of protocol (SHAP=0.06) and PCOS (SHAP=0.01) were far less significant to the prediction outcomes. Our model suggests that AMH is over 8 times more important for the prediction of number of retrieved MII than age. The algorithm found that 300 IU dose of gonadotropin administered on days 4-7 of stimulation actually lowered the number of retrieved MII oocytes. Limitations, reasons for caution Only 5% of study patients had PCOS, which might contribute to low importance of this factor in the model. Wider implications of the findings Research shows that AMH and AFC are the most impactful factors in MII oocytes prediction by a large margin and should be considered in a first place by gynecologists while choosing gonadotropin dose. To grow more MII oocytes, clinicians should use higher gonadotropin dose in days 1-3. Trial registration number Not applicable