BACKGROUND
Recommendation systems (RS) have been widely used in the field of nutrition to promote the nutritional self-management, but few NRSs have been widely adopted due to various reasons. Limited studies have reviewed the RSs in food, with some methodological flaws including limited databases searches, high heterogeneity among included studies and rapidly evolving nature of evidence.
OBJECTIVE
We conducted a scoping review to summarize currently available recommendation systems applied in nutrition (NRS) including published articles, patents and application software and explore the potential gaps between development and implementation.
METHODS
We conducted a comprehensive search of seven bibliographic databases, two patent databases, four mobile apps store and three websites engines for this scoping reviews. Data extraction was conducted by four reviewers, a pilot study was performed before formal extraction, and the interrater agreement percentage needed to be >75%. Discrepancies were resolved by consensus or the involvement of a third reviewer. Frequency count and narrative summaries were performed for each study.
RESULTS
A total of 877 NRSs were included and half of them were released after 2022 (n=423, 48.2%) and 155 (17.7%) were from China. The most users were overweight or obese population (n=152, 17.3%), the primary inputs being self-reported data on nutritional status, diet, and exercise (the same n=157, 17.9% for every one), the primary output being nutrition plans (n=254, 29.0%), and the main audience being general population (n=244, 27.8%). Of 49 studies published in journals or essays, a few researchers from the nutrition filed were reported (n=4, 3.6%), the primary data were from public survey (n=22, 46.8%). Forty studies reported the evaluation stage, with incomplete processes and the lack of nutritional outcome. Of the 18 artificial intelligence technologies used in the studies, four could automatically update systems by themselves, and two were technologies proposed in the last decade. In addition, three recommendation algorithms were identified, only one was the latest knowledge-based algorithm that can improve precise matching.
CONCLUSIONS
While NRSs has primarily focused on the general population, there is a growing demand for professional NRSs tailored to special populations that incorporate dynamic updates and enhanced individual identification. Standardized evaluation of NRSs based on their technical performance and clinical impact can effectively support their public application in future.
CLINICALTRIAL
The protocol was registered on the Open Science Framework (https://doi.org/10.17605/OSF.IO/VF7NB)