Hybrid AI for Clinical Prediction: Combining LLM-Derived Features with ML for Explainable Diabetes Risk Assessment

The increasing global prevalence of diabetes requires the development of accurate and early predictive models. Traditional machine learning (ML) approaches often struggle with interpretability and the ability to leverage rich and unstructured clinical insights. This study introduces a novel hybrid AI framework that combines a traditional ML model (Bagging Classifier) with features derived from a large language model (LLM), specifically the Gemini model, for improved prediction of diabetes. We used LLM to generate high-level interpretable features (e.g. sentiment scores, keyword flags, and semantic embeddings) from patient profiles constructed using structured data, which were then integrated into a meta-learner model. Although a proprietary LLM was used in this study, the methodology was designed to be model-agnostic and reproducible using open-source software. Our framework was evaluated on a combined dataset of PIMA Indian Diabetes and CDC BRFSS2015 data, demonstrating superior performance, achieving a recall of 0.83 ± 0.009 (95% CI: [0.821, 0.839]) and an accuracy of 0.90 ± 0.019 (95% CI: [0.881, 0.919]) in unseen test data. Beyond performance, the hybrid model offers enhanced explain-ability through SHapley Additive exPlanations (SHAP) values, revealing key feature contributions from both traditional and LLM derived sources. We also conducted a comprehensive fairness analysis to highlight the model's equitable performance across demographic groups.