(551) Predictive Risk Factors for Hypogonadism Using Interpretable Machine Learning Algorithms Based on Shap: A Large National Cohort Study

Introduction Hypogonadism is a prevalent condition associated with numerous adverse health outcomes. While many risk factors have been identified, conventional linear statistical models are often inadequate for capturing the complex non-linear relationships and interactions among them. Objective This study aimed to develop a high-precision risk prediction model for hypogonadism by leveraging interpretable machine learning (ML) methods. Methods We utilized data from 2,524 adult male participants from the 2011–2016 National Health and Nutrition Examination Survey. We developed and compared six ML models. We employed SHAP (Shapley Additive exPlanations) to quantify individual feature contributions and Partial Dependence Plots (PDP) to visualize the marginal effects of features on the predicted outcome. Results The XGBoost model demonstrated the best predictive performance, achieving an Area Under the Curve of 0.856, an accuracy of 0.809, a sensitivity of 0.729, and a specificity of 0.828. SHAP analysis identified SHBG, age, Triglyceride-glucose body mass index (TyG-BMI), TyG-waist-to-height index, and hemoglobin as the five most important predictors. PDP analysis revealed complex non-linear relationships, including distinct plateau effects; Critically, we identified a significant "buffering effect", effectively neutralizing the adverse impact of other factors like advanced age or metabolic dysfunction. The model was deployed as an interactive web. Conclusions This study successfully constructed a highly accurate and interpretable ML model for predicting hypogonadism risk. The quantitative analysis of risk plateaus and the novel discovery of SHBG's buffering effect provide new evidence for the complex etiology of hypogonadism, laying a solid foundation for future precision prevention and personalized treatment strategies. Disclosure No