AI-Driven Personalized Nutrition: A Machine Learning Framework for Dietary Recommendations Based on Genetic, Metabolic, And Lifestyle Data

Artificial intelligence personalizes nutrition; however, research remains fragmented, lacking integrated frameworks linking multimodal data, prediction, recommendations, and validation. This systematic literature review analyses primary human evidence on machine-learning-driven dietary personalization and develops a decision-support framework to distinguish prediction from recommendation. The literature search was conducted using Scopus, Web of Science plus, IEEE Xplore, PubMed/MEDLINE-indexed literature and Europe PMC and, via citation linkage, obtained a verified screening database of 53 unique records. After de-duplication, 53 unique records were identified. Twenty-one were excluded during title/abstract screening, leaving 32 records for further assessment. Five additional records were excluded before full-text retrieval, 27 reports were sought for retrieval, and six could not be retrieved. The remaining 21 full-text reports were assessed for eligibility; one was excluded because it was not directly related to the review question, yielding 20 studies for qualitative synthesis. The included studies comprised externally validated prediction studies, randomized dietary intervention trials, digital-twin interventions, recommender systems, microbiome-guided diets, metabolomics, and multi-omics nutrigenetics. Multimodal models had better predictive performance with regard to postprandial glycemia, triglyceride response, weight-loss response and metabolite response as compared with simpler dietary variables in selected settings. However, this predictive performance in itself did not constitute clinical benefit. The evidence was less consistent for clinical effectiveness. A number of randomized trials demonstrated an advantage with personalized strategies regarding glycemic control. In contrast, other ones failed to prove a significant advantage in terms of weight loss, LDL cholesterol, and severity of primary irritable bowel syndrome. These results suggest that the effectiveness of predictive accuracy should be considered as two distinct evidence questions. The review provides an evidence-informed, 5-layer decision-support architecture to define a clear separation between prediction, recommendation, validation, and governance of integrated multimodal nutrition evidence and clinical decision support to overcome fragmentation across genetic, metabolic, lifestyle, microbiome, and sensor-derived domains.