**Background:** Recommender systems are AI-based tools that help users navigate large information spaces by providing personalized suggestions. While these systems have been successfully applied in e-commerce, e-learning, and e-tourism, their application in e-health—particularly for food recommendations for diabetic patients—remains underdeveloped. Diabetes is a global health emergency, with 537 million adults living with the condition in 2021, costing an estimated US$827 billion per year. Unhealthy diets are a major risk factor, and personalized nutrition approaches could significantly improve patient outcomes. Despite growing interest in health-aware food recommender systems, no comprehensive analysis had specifically focused on the diabetes domain. This systematic review, guided by the PRISMA 2020 framework, addresses three research questions: (RQ1) What are the current state-of-the-art research works on food recommender systems for diabetic patients? (RQ2) What computational methods and tools are used? (RQ3) What evaluation frameworks measure effectiveness?
**Methods:** A literature search was conducted in the first week of December 2022 across three electronic databases: Web of Science Core Collection, Scopus, and PubMed (including Medline). Studies from January 2010 to November 2022 were considered. The search strategy used three query strings: (1) "food" AND ("recommender systems" OR "recommendation systems") AND ("diabetics" OR "diabetes"); (2) "intelligent" AND "nutrition" AND ("diabetics" OR "diabetes"); (3) "food" AND "personalized" AND "systems" AND ("diabetics" OR "diabetes"). Inclusion criteria required papers to focus on food recommendations and present novel methods or systems incorporating personalization. Exclusion criteria removed papers without computational methods, those not considering diabetes as a relevant feature, and abstracts or small reports lacking detail. Two reviewers screened titles, abstracts, and keywords; full-text analysis was performed by two additional reviewers, with disagreements resolved through discussion. Quality evaluation used a 7-element checklist adapted from Ataei and Litchfield (2022), covering minimum threshold quality, rigour, credibility, and relevance. Studies required at least 75% positive responses to be included.
**Key Results:** The initial search yielded 967 records (72 from Web of Science, 129 from Scopus, 766 from PubMed). After removing 302 duplicates, 665 records were screened by title and abstract, yielding 95 papers for eligibility assessment. Of these, 91 were retrievable; 12 were excluded for lacking computational methods, 18 for not focusing on diabetics, and 20 for being small reports. After quality evaluation, 34 papers were included in the final analysis. The papers were grouped into a four-category taxonomy based on term co-occurrence analysis of abstracts: (1) Semantic-based approaches (10 papers) using ontologies and inference mechanisms like Jena inference engine and case-based reasoning; (2) Optimization-based approaches (8 papers) using integer programming, population-based metaheuristics, ant colony optimization, and constraint satisfaction; (3) Rule-based and classification approaches (10 papers) using IF-THEN rules, neuro-fuzzy inference, AHP, clustering, and decision trees; (4) Interaction-based approaches (6 papers) using dynamic questionnaires, chatbots, hardware sensors, and multi-agent architectures. Key weaknesses across all groups included limited evaluation frameworks, lack of standardized datasets, insufficient incorporation of diabetes-specific knowledge, and poor reproducibility. Only a few studies (e.g., Wang et al., Nag et al.) presented complete research cycles with public datasets and appropriate evaluation protocols.
**Clinical Implications:** This review reveals that current food recommender systems for diabetic patients lack the rigor needed for clinical deployment. Most systems fail to adequately incorporate established medical guidelines for diabetes nutrition, such as low-carbohydrate, Mediterranean, and high-protein diets shown to improve cardiovascular risk markers. The limited evaluation approaches—often restricted to case studies or synthetic data—mean that clinical effectiveness and patient safety remain unproven. The review identifies critical future directions: developing a consolidated research framework integrating optimization and interaction-based approaches; better formalizing the research problem (individual foods vs. menus vs. recipes); exploiting recent medical knowledge on diabetes nutrition guidelines; incorporating user preferences more intensively; using fuzzy tools for uncertainty management; and developing explainable AI approaches as required by EU ethics guidelines. Without addressing these gaps, the potential for these systems to improve outcomes for the 537 million adults living with diabetes will remain unrealized.