The construction of machine learning-based predictive models for progression to premature ovarian insufficiency within three years in patients with diminished ovarian reserve

Abstract Background Premature ovarian insufficiency (POI) leads to infertility in women before the age of 40 and significantly diminishes patients' quality of life. Diminished ovarian reserve (DOR) represents an earlier stage of ovarian aging compared to POI. However, clinical tools for prediction of this progression are currently lacking. Accordingly, this study sought to construct an interpretable machine learning-based prediction model to identify DOR patients at high risk of progressing to POI within three years. Methods We enrolled 312 patients diagnosed with DOR at the First Hospital of Hunan University of Chinese Medicine between January 2014 and January 2024. After addressing missing values, the dataset was partitioned into training (70%) and validation (30%) cohorts by using stratified random sampling. Multicollinear variables were first removed via correlation analysis. Then we used LASSO and the Boruta, taking the intersection of variables selected by both methods. Seven machine learning algorithms were used to develop models. To assess model performance, we used the area under the curve (AUC), accuracy, sensitivity, specificity, calibration curves, and decision curve analysis. The SHapley Additive exPlanations (SHAP) method was then applied to provide visual interpretations of the optimal one. Results A total of 312 patients with DOR were included, of whom 153 (49.0%) progressed to POI within the three-year follow-up period. After comparing seven different models, the Random Forest model demonstrated the best performance. SHAP analysis revealed follicle-stimulating hormone (FSH) as the primary risk factor and identified seven key features contributing to POI progression, along with the direction of their effects. Conclusion The interpretable Random Forest model established in this study enables prediction of three-year progression risk from DOR to POI. By facilitating identification of high-risk patients, this model may provide useful risk estimates to guide clinical decision making and patient counseling.