FL-EHDS: A Privacy-Preserving Multimodal Federated Learning Framework for the European Health Data Space

The European Health Data Space (EHDS), established by Regulation (EU) 2025/327, mandates cross-border health data analytics while preserving citizen privacy. We present FL-EHDS, a three-layer compliance framework integrating EHDS governance (Health Data Access Bodies, data permits, Article 71 opt-out) with federated learning orchestration (17 algorithms including ICML/ICLR 2024±2025 advances, differential privacy, secure aggregation) and data holder components (FHIR R4 preprocessing). Multimodal validation across 6,004+ experiments on tabular clinical, ECG, and medical imaging datasets demonstrates that personalization architecture is the critical deployment decision on compact clinical models, with the effect becoming method- and regime-dependent on complex imaging architectures: only methods maintaining separate local models (Ditto, HPFL) differentiate from FedAvg on compact clinical models, producing up to 26.8pp accuracy gainsÐan advantage confirmed as hyperparameter-insensitive (≤1.44pp across 100× λ variation), model-architecture invariant (within the tabular family: MLP and TabNet, 2.9K±701K parameters), and persistent under data quality degradation and compound EHDS stress (+9.6pp mean, 81% of conditions). EHDS governance adds <1.1% per-round overhead. Privacy at ε=10 costs<2pp; cross-border heterogeneous per-client privacy budgets are viable (−0.9pp); full EHDS compliance (simultaneous data minimization, opt-out, and DP at ε=10) costs Ditto only −0.7pp (p<0.001). On imaging (ResNet-18), personalization is methoddependent: Ditto excels (+25.5pp Skin Cancer) while HPFL fails (−18.2pp Chest X-ray). A novel Diagnostic Equity Index reveals that accuracy masks severe per-class disparities correctable only by personalized algorithms. The open-source implementation provides deployment guidance for the 2029 secondary use deadline.