AI-Driven Diabetes Risk Assessment and Personalized Health Communication: A Review of Machine Learning, Explainable AI, and Generative AI
Main Article Content
Abstract
Generative Artificial Intelligence (GenAI), Large Language Models (LLMs), and Retrieval-Augmented Generation (RAG). However, these technologies have largely evolved as separate research directions, creating a gap between accurate risk prediction and understandable, trustworthy health communication. This critical review examines the progression of AI-based approaches for diabetes risk assessment, with particular emphasis on ML prediction, XAI, GenAI/LLMs, RAG, and intelligent healthcare assistants. The review synthesizes existing evidence to examine how ML models support risk estimation, how SHAP and LIME improve model interpretability, and how GenAI and LLMs enable natural-language and personalized healthcare communication. It further considers RAG as a mechanism for grounding generated responses in external knowledge sources and reducing dependence on internally generated information. The critical synthesis identifies persistent gaps in the integration of prediction, explanation, evidence grounding, personalization, and safety. Existing systems frequently provide either predictive outputs, technical explanations, or general conversational support rather than combining these capabilities within a unified diabetes risk-awareness framework. Based on these findings, an integrated conceptual framework is proposed linking ML prediction, XAI-based risk interpretation, RAG, GenAI/LLM-based communication, personalization, and safety validation. The review highlights the need for future empirical validation, standardized evaluation, privacy-aware personalization, and patient-centred assessment of integrated AI healthcare assistants.