Development and Validation of an Interpretable Machine Learning Model for Predicting Diabetes Mellitus in Hospitalized Patients

OBJECTIVE: Early detection of diabetes mellitus (DM) is critical for preventing disease progression. This study aimed to develop an online explainable risk prediction model to assess individual susceptibility to DM. METHODS: A total of 356 hospital patients with 46 demographic, behavioral, and laboratory indicators were included. Recursive feature elimination based on Random Forest was used for feature screening. Six machine learning models Extremely Randomized Trees (ET), Extreme Gradient Boosting, Light Gradient Boosting Machine, Logistic Regression, Random Forest, and Multi-Layer Perceptron were constructed, with hyperparameters optimized by grid search and fivefold cross-validation to ensure robustness. Model performance was evaluated using receiver operating characteristic curves, decision curve analysis, and calibration curves. SHapley Additive exPlanations was used to interpret predictor contributions, and a personalized online platform was developed. RESULTS: Eight key features (HbA1c, CHA2DS2VASC, FT4, ALT, Hb, DiastolicBP, AST, and BMI) were identified. The ET model demonstrated optimal performance, achieving a test set accuracy of 92.21%, sensitivity of 71.43%, F1-score of 76.92%, and area under the curve of 0.912-0.996. It exhibited excellent discrimination, calibration, and clinical utility across diverse validation populations. CONCLUSIONS: This study developed a clinically applicable, explainable online DM risk prediction platform. It enables efficient individual risk assessment for healthcare providers, facilitating timely interventions to improve DM prevention and management.