Applications of artificial intelligence in personalized nutrition interventions

Abstract

Unhealthy dietary patterns remain a major modifiable contributor to the global burden of obesity, type 2 diabetes mellitus (T2DM), cardiovascular disease, and cancer. However, individuals can respond markedly differently to identical foods, limiting the effectiveness of one-size-fits-all dietary guidance. This review examines the application of artificial intelligence (AI) to personalized nutrition interventions across three interconnected layers. At the biological level, inter-individual variability can be characterized by three mechanistic dimensions: nutrigenomic variation, which influences nutrient metabolism and disease susceptibility; metabolomic biomarkers, which objectively capture dietary exposure and early metabolic alterations; and the gut microbiome, which bidirectionally mediates diet–host interactions. At the technological level, AI integrates these dimensions into multimodal digital profiles. Machine learning models predict individual postprandial glycemic and lipemic responses, with reported correlations of up to r = 0.77 in large cohorts; deep learning and computer vision automate dietary assessment; natural language processing and large language models (LLMs) support nutrition counseling; knowledge graphs enable constraint-aware food recommendations; and digital-twin and adaptive closed-loop systems incorporating continuous glucose monitoring and wearable-device data support real-time nutritional interventions. However, the strength of clinical evidence varies across disease domains. Glycemic management in T2DM and prediabetes currently has the most mature randomized controlled trial evidence: an AI-enabled personalized intervention combining algorithmic dietary personalization with CGM-linked feedback improved glycemic endpoints compared with a Mediterranean diet, while a fully automated AI-delivered diabetes prevention program achieved outcomes non-inferior to those of human coaching. In obesity, cardiovascular disease, oncology, and gastrointestinal disorders, the evidence remains mixed or preliminary. Clinical translation also faces substantial challenges, including biased and inconsistently standardized dietary and multi-omics data; the underrepresentation of non-European populations, which limits model generalizability and raises concerns about health equity; limited model explainability; privacy and data-governance challenges; immature regulatory pathways for adaptive software as a medical device; and specific risks associated with LLMs, such as hallucinated or guideline-discordant dietary advice. AI-enabled personalized nutrition is therefore progressing from proof of concept toward clinical translation. Realizing its potential equitably will require multi-ethnic nutrition–omics cohorts, privacy-preserving multicenter collaboration, dietitian-led human–AI collaboration, and lifecycle safety governance for generative models.