Pdiapredict-Explainer: An Interpretable AI System for Personalized Diabetes Risk Management

Abstract

This study presents an interpretable LightGBM model designed to predict diabetes risk, developed using the BRFSS2015 dataset, which includes more than 250,000 records. To make which is called the model more understandable and tackle the common "black box" issue, we incorporated feature engineering—such as creating a Metabolic. Risk score—and applied SMOTE to handle class imbalance. The resulting model achieved an AUC-ROC of 0.8264 and was particularly effective at identifying low-risk individuals, with a 98% recall rate, which can help simplify screening efforts. Using SHAP (SHapley Additive exPlanations), we visualized how different factors contribute to each person's risk, including waterfall and decision plots that break down individual cases—like showing how metabolic dysfunction might account for roughly 31% of a person's risk. This approach supports personalized intervention strategies aligned with ADA guidelines. However, the model's ability to catch high-risk cases was limited, with only an 18% recall, emphasizing the difficulty of detecting complex risk profiles without biochemical data like HbA1c. Moving forward, integrating such markers and tracking patients over time with electronic health records could improve performance. Overall, this framework illustrates how explainable AI can build trust in preventive diabetes strategies by making predictions transparent and clinically relevant.