Overview
The present study addresses the need for validating federated learning frameworks using real clinical data, specifically focusing on the FedCVR framework. This validation represents a crucial step between proof-of-concept demonstrations in synthetic environments and the deployment of such systems in real multicenter healthcare settings. The research provides empirical evidence for the FedCVR framework's clinical viability in multicenter contexts.
Research Context
Earlier architectural work by the same authors (Tertulino and Alencar, 2026) utilized a synthetic six-feature benchmark. This previous study found that server-side adaptive optimization acted as a temporal denoiser for Differential Privacy (DP) noise, addressing a challenge identified in a prior pipeline work (Tertulino, 2025). The earlier research, however, relied on synthetically generated data and explicitly identified real-world validation as a priority for future investigations. The current study was designed to bridge this gap by conducting validation on genuine clinical datasets.
Approach
The research validated the FedCVR framework using five publicly available real cardiovascular datasets. These datasets included Framingham, Cleveland, Hungarian, Switzerland, and Long Beach VA. To ensure consistency for the federated learning setup, these datasets were harmonized to conform to the 13-attribute UCI Heart Disease schema. The study configured these harmonized datasets into a heterogeneous federated scenario. Cross-validation was performed using a leave-one-institution-out methodology. The performance of FedCVR was evaluated against standard FedAvg.
Findings
- The FedCVR framework maintained its adaptive advantage on real clinical data.
- Under an operational privacy budget (noise multiplier = 0.8, privacy budget $\epsilon \approx 4.2$), FedCVR achieved an F1-Score of 79.2%.
- Under the same operational privacy budget, FedCVR achieved an Area Under the Curve (AUC) of 0.96.
- FedCVR statistically outperformed standard FedAvg across all evaluated metrics. This superiority was confirmed by paired t-tests, where all p-values were 0.003, indicating significance under the Bonferroni-corrected threshold.
- The measured privacy cost on real data corroborated the graceful degradation pattern previously observed in synthetic experiments.
Why This Matters
The empirical validation of the FedCVR framework on real cardiovascular datasets provides evidence of its clinical viability within genuine multicenter healthcare environments. This suggests that the framework can maintain its performance advantages while operating under specified differential privacy constraints, which is crucial for secure and collaborative analysis of sensitive clinical data across multiple institutions.