Federated learning in a regulator-audited secure processing environment: a multi-hospital deployment study

The study presents the implementation of federated learning in a certified secure processing environment at Helsinki University Hospital (HUS) in Finland. Federated learning enables collaborative model development without transferring patient data between institutions. The system was deployed across three university hospitals (Helsinki, Turku, Tampere) and underwent a differential security assessment that identified five findings (one high-severity, two medium-severity, and two low-severity), all of which were remediated. A federated DeepSurv model for predicting acute myeloid leukaemia prognosis was successfully trained without centralizing patient data, achieving 100% parameter merge success. The federated model demonstrated superior risk stratification compared to locally trained models, with higher log-hazard ratios (4.32 vs 1.79 at HUS, 7.20 vs 3.19 at TAYS, and 5.90 vs 2.91 at TYKS). The study demonstrates a practical pathway for operationalizing federated learning within regulated healthcare environments.