Machine Learning Interpretability in Diabetes Risk Assessment: A SHAP Analysis

Diabetes continues to be a complicated and prevalent metabolic illness, providing a serious burden to public health. While machine learning approaches like extreme gradient boosting (XGBoost) provide intriguing options for diabetes prediction, their 'black-box' nature typically limits clinical interpretability. To overcome this gap, our work applied SHapley Additive exPlanations (SHAP) to give insights into the XGBoost model's predictions. The dataset utilized in this research comprised of 253,680 patients and contained 21 parameters, such as General Health Status, High Blood Pressure Status, Age, and Body Mass Index. After feature selection using Recursive Feature Elimination (RFE), 15 important characteristics were discovered. In the test set, the XGBoost model obtained an accuracy of 86.6%, precision of 54.1%, recall of 17.0%, and an F1-score of 25.9% for the Original dataset. For the RFE dataset, the model displayed an accuracy of 86.6%, precision of 54.9%, recall of 16.5%, and an F1-score of 25.3%. SHAP analysis found that General Health Status, High Blood Pressure Status, Age, and Body Mass Index were the most important characteristics in both the Original and RFE datasets. This work provides as a platform for transparent and clinically applicable predictive modeling, assisting in early diabetes identification and preventive healthcare.