Federated Learning for Multi-Institutional Healthcare AI: Privacy, Generalizability, and Governance

摘要

Background: The adoption of federated learning in cross-hospital research, rare conditions, and distributed model development creates opportunities to improve clinical services, but technical performance alone does not establish patient benefit. Objective: To critically examine applications, evidence, safety, equity, and implementation of federated learning. Approach: This narrative review develops an analytical synthesis of established clinical AI principles and reporting guidance; it does not report a systematic search, a new dataset, or an empirical trial. Findings: Key opportunities include distributed training, external evaluation, and privacy-aware collaboration. Important limitations include non-identical populations, variable labels, and communication failures, as well as model leakage, participation imbalance, and governance failures. Responsible adoption requires explicit data agreements, security testing, and withdrawal procedures, transparent reporting, and evaluation of cross-site discrimination, calibration, robustness, and clinical utility. Conclusion: The clinical value of federated learning depends on fit with the care pathway, prospective evaluation, accountable human oversight, and sustained monitoring after deployment.