Generative Artificial Intelligence for Diabetes Mellitus Prediction and Intelligent Clinical Decision Support

Abstract

Diabetes mellitus is a major chronic metabolic disorder characterized by persistent dysregulation of blood glucose and progressive cardiovascular, renal, neurological, and ophthalmic complications. The global prevalence of diabetes has increased substantially, reaching approximately 14% among adults in 2022, with nearly 830 million people living with the disease. The rapid expansion of electronic health records, continuous glucose monitoring, laboratory databases, medical imaging, wearable technologies, and patient-generated health data has created opportunities for artificial intelligence (AI)-based diabetes prediction and clinical decision support. Conventional machine learning models demonstrate potential for diabetes detection and complication prediction; however, challenges remain regarding heterogeneous datasets, external validation, calibration, dataset shift, interpretability, and clinical implementation. This study proposes a systematic framework integrating generative AI, particularly large language models (LLMs), with validated predictive models for diabetes risk assessment and intelligent clinical decision support. A secondary evidence-synthesis approach was used, integrating findings from systematic reviews, clinical trials, diabetes technology research, and recent LLM literature. Evidence indicates that only 36% of evaluated machine learning models for diabetes complications demonstrated clearly useful discrimination. Generative AI has shown promise in treatment support, although performance decreases with increasing clinical complexity. ChatGPT-4 achieved 69.2–84.6% agreement with clinician prescribing in selected monotherapy cases but only 47.1–49.8% for dual- or triple-therapy cases. The findings support using generative AI as an orchestration layer for prediction, evidence retrieval, contextualization, explanation, documentation, and workflow support rather than as a replacement for validated prediction or clinical judgment. Safe implementation requires calibrated models, authoritative retrieval, uncertainty representation, auditability, privacy, fairness assessment, human oversight, and prospective clinical validation.