Federated learning for clinical outcome prediction from structured and temporal health data: A systematic review of performance, heterogeneity, privacy, and clinical translation

Abstract

Retrospective registration of a PRISMA 2020 systematic review of empirical federated learning (FL) studies that predict individual patient clinical outcomes from structured electronic health records, clinical time series, sensor and text data. The review addresses six research questions: FL algorithms and performance relative to centralized (pooled) and local-only baselines; prediction tasks, temporal structure, datasets and architectures; handling and measured effect of non-IID heterogeneity; privacy mechanisms beyond the base protocol and their utility and resource costs; validation and deployment readiness; and recurring methodological limitations and risk of bias (PROBAST). Five sources were searched on 13 August 2026 (PubMed/MEDLINE, IEEE Xplore, ACM Digital Library, ACL Anthology, Google Scholar). Synthesis follows SWiM, using within-study direction-of-effect vote counting, and certainty is rated with a GRADE-informed scheme for prognostic evidence. This registration was made after screening and data extraction had been completed. It records the protocol that was fixed before full-text screening and states which analyses were added after that point.