Application of Artificial Intelligence and Large Language Models in Clinical Decision-Making for Diabetes

Abstract

The global prevalence of Type 2 Diabetes Mellitus (T2DM) and the complexity of its management pose significant challenges to traditional healthcare models. In recent years, the rapid advancement of Artificial Intelligence (AI) technologies—particularly Machine Learning (ML), Deep Learning (DL), and Large Language Models (LLMs)—has provided revolutionary tools for developing intelligent, precise, and personalized Clinical Decision Support Systems (CDSS). This review systematically maps the application landscape of AI and LLMs across the full-cycle management of T2DM, covering the complete chain from risk prediction and precise diagnosis to individualized treatment and long-term prognosis management. In risk prediction, AI models significantly improve early identification of progression from prediabetes to T2DM, as well as diabetic nephropathy, cardiovascular events, and other complications by integrating multidimensional data. At the diagnostic level, in addition to differential typing based on electronic health records, emerging studies have digitized traditional Chinese medicine tongue images and combined them with biomarkers such as oral-gut microbiota through multimodal machine learning, opening new pathways for non-invasive and objective auxiliary diagnosis. In treatment decision support, AI not only recommends glucose-lowering medications and optimizes insulin dosages but also predicts individualized postprandial glucose responses using continuous glucose monitoring data, offering possibilities for dynamic behavioral interventions. LLMs, exemplified by ChatGPT, demonstrate strong potential in interpreting clinical text, simulating doctor-patient communication, and generating preliminary management plans through their powerful natural language processing capabilities, although the reliability and safety of their independent decision-making still require cautious evaluation. Nevertheless, a substantial gap remains between technological advancement and clinical translation. This review further analyzes the core challenges currently faced, including data quality and algorithmic bias, the “black-box” nature of models and insufficient explainability, barriers to integration into clinical workflows, and profound medical ethical issues such as equity, privacy protection, and the risk of marginalizing human decision-making agency. Finally, the paper outlines future directions for the next generation of intelligent CDSS that integrate multimodal data, federated learning, causal inference, and human-centered design principles, emphasizing the need to strike a balance between technological innovation and ethical governance in order to build a safe, effective, equitable, and trustworthy AI-empowered paradigm for diabetes management.