Using an Interpretable Machine Learning Model to Predict Corifollitropin Alfa Protocol

Abstract

Background: To demonstrate an interpretable machine learning (ML) model for a clinical prediction of corifollitropin alfa protocol. Methods: The retrospective study involved 1,221 cycles from 1,180 patients undergoing corifollitropin alfa protocol with oocyte retrieval events from a single in vitro fertilization (IVF) center. The ML models were assigned to the following tasks, which are the dosage of corifollitropin alfa, trigger type, the dosage of recombinant FSH (rFSH), the dosage of recombinant LH (rLH), the duration between the follow-up visit (FUV), and oocyte retrieval. Interpretable SHapley Additive exPlanations (SHAP) were selected to analyze the input features. The ranking of the prediction powers from each input feature reveals the level of contribution to the model. Result(s): Two series of interpretable ML models were developed to predict classification tasks and regression tasks. The areas under ROC (AUC) for predicting the dosage of corifollitropin alfa and trigger type were 0.933 ([Formula: see text] CI 0.907–0.958) and 0.891 ([Formula: see text] CI 0.864–0.918), while accuracies were 0.944 and 0.904. The mean absolute errors (MAEs) that predict the dosage of rFSH, the dosage of rLH, and the duration between the FUV and oocyte retrieval were 97.08 IU (rFSH), 105.61 IU (rLH), and 0.45 days. Conclusions: The study demonstrates a set of interpretable ML models predicting tasks involved in corifollitropin alfa protocol. The potential for the clinical application is to provide consistency in corifollitropin protocol adjustments.

Keywords

• Artificial Intelligence
• In Vitro Fertilization
• Machine Learning
• Controlled Ovarian Stimulation
• Interpretable Approach
• Gradient Boosting Decision Tree