Centralized pooling and federated learning for Canadian patient-level data sharing in multicenter medical AI: A...

Algorithms that support screening, triage, and treatment decisions depend on training data drawn from patient populations. Limited access to patient-level records across institutions and jurisdictions can reduce representation and contribute to uneven model performance across populations. Canada's federated health system, where provinces and territories manage separate datasets and privacy regimes, limits multicenter medical AI research. We conducted a scoping review to map how Canadian researchers share patient-level data in multicenter medical AI collaborations. We searched PubMed, IEEE Xplore, ACM Digital Library, Scopus, and Web of Science from 2018 to February 2025 and implemented a human-in-the-loop large language model process to support screening and extraction, with reviewer validation. Among 3100 included studies, 160 reported multicenter patient-level data collection. Centralized pooling dominated this subset, with 95% of studies using centralized storage and 5% (n = 8) reporting decentralized approaches, including federated learning, sequential model transfer, and distributed feature sharing. Governance requirements were frequently described as multi-site and sequential, and 81.8% of multicenter collaborations reported parallel ethics approvals from three or more institutional review boards. Only one decentralized collaboration operated entirely within Canada. International partnerships comprised 80% of multicenter studies, and many cohorts included non-Canadian sites or non-Canadian data. Our findings support adoption of distributed model development protocols and interoperable governance that limit central pooling while enabling consistent training, validation, and reporting across sites, as only 1 of 160 multicenter studies reported a decentralized approach with Canadian patient data only.