Federated Clinical Learning Cooperative (FCLC)

Background: Developing robust clinical artificial intelligence (AI) models requires large, diverse datasets that individual institutions cannot provide due to privacy regulations (GDPR, HIPAA) and institutional risk aversion. Existing federated learning (FL) platforms lack validated combinations of differential privacy, Byzantine robustness, fair contribution attribution, and secure aggregation. Objective: We present the Federated Clinical Learning Cooperative (FCLC), an open-source platform enabling multi-institutional clinical AI development without raw data leaving participating sites. Methods: FCLC implements a data preprocessing pipeline (Layers 1-3: direct identifier removal, quasi-identifier generalization, k-anonymity with k≥5) combined with differential privacy (Layer 4: DP-SGD, ε=2.0/round, δ=10⁻⁵, Rényi accountant α=4.0) and secure aggregation (Layer 5: SecAgg+ via CommonHealth). Validation used MIMIC-IV (N=12,543, 30-day readmission) and eICU-CRD (N=8,420, sepsis mortality) across IID and non-IID partitions (Dirichlet α ∈ {∞, 1.0, 0.5, 0.1, 0.01}) with logistic regression and multilayer perceptron architectures. Results: On MIMIC-IV, FCLC achieved AUC=0.758 [95% CI: 0.739–0.777] compared to centralized oracle 0.789 (Δ = −3.9%). Under severe non-IID conditions (Dirichlet α=0.1, EMD=0.31), FCLC preserved AUC=0.748 while FedAvg degraded to 0.694 (p_adj=0.003). Membership inference attack AUC with DP was 0.52±0.03 (indistinguishable from chance, p=0.31 vs. 0.50). At ΔAUC=0.03 — the minimum clinically important difference (MCID) corresponding to preventing approximately one readmission per 100 patients (NNT≈100) — the study had power >0.99. Conclusions: FCLC provides a validated, regulation-compliant infrastructure for federated clinical AI with full cryptographic privacy guarantees suitable for mutual-distrust deployments. The platform is open-source (Apache 2.0) and fully reproducible via Docker.