Federated learning in multicenter digestive disease research: A scoping review of privacy-preserving technology applicability (Preprint)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

BACKGROUND Federated learning (FL) offers a potential way to balance privacy protection with data use in multicenter research on digestive diseases by keeping data local while models are trained collaboratively across centers. However, the application landscape of FL in the digestive field and the extent to which privacy-preserving techniques have been implemented remain poorly understood. OBJECTIVE To map the evidence landscape of FL in multicenter digestive disease research and analyze the current use and suitability of four privacy-preserving techniques: differential privacy (DP), secure aggregation (SA), homomorphic encryption (HE), and secure multiparty computation (MPC). METHODS We conducted a scoping review using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews checklist across eight databases (PubMed, Europe PMC, Semantic Scholar, arXiv, Crossref, Embase, Web of Science, and CNKI). After deduplication by DOI and normalized title and screening using a unified taxonomy, 77 original studies were included, and privacy-preserving techniques were assessed at the full-text and abstract levels. RESULTS Among the 77 included studies, 61 (79.2%) were related to imaging/endoscopy/pathology, 14 (18.2%) to structured tabular data, and 2 (2.6%) to genomics. The diseases were predominantly hepatic (22 studies), colorectal (17), and pancreatic (12), and only 27 studies (35.1%) used real-world multicenter data. Moreover, seven studies (9.1%) implemented an algorithmic privacy-preserving technique (DP, SA, HE, or MPC), whereas 70 (90.9%) used none. “Deployment security” was systematically conflated with “algorithmic privacy”. Regarding governance arrangements, 15 studies (19.5%) explicitly reported ethics approval; however, the reporting rates for data-sharing agreements, aggregation-server ownership, intellectual-property arrangements, and data-governance frameworks were significantly low—only five studies (6.5%) reported at least one of these four, and 34 (44.2%) reported none of the five fields. These findings indicate substantial gaps both in algorithm-level privacy protection and in the reporting of institutional governance arrangements. CONCLUSIONS FL in the digestive field is still in its early stages. Privacy requirements have not translated into engineering practice. Structured tabular data remain a major methodological gap, and institutional governance arrangements (data governance, data-sharing agreements, aggregation-server ownership, and intellectual property) are almost entirely unreported.