Edge-Driven Federated Learning for Healthcare Data Privacy

Thesis Abstract

The increasing digitization of health records and the proliferation of edge devices in clinical environments heighten the risk of data breaches and compromise patient privacy, while traditional centralized learning paradigms raise concerns about data sovereignty and latency. This study addresses the problem of enabling robust and privacy-preserving machine learning for healthcare while reducing transmission costs and respecting regulatory constraints through edge-driven federated learning. The aim is to design, implement, and evaluate an edge-aware federated learning framework that optimizes privacy guarantees, communication efficiency, and model performance for diverse healthcare tasks executed on distributed hospital and device-level data. Specific objectives are to (1) quantify the privacy-utility trade-offs of edge-based federated learning under differential privacy and secure aggregation, (2) develop adaptive client selection and compression strategies to minimize communication overhead without sacrificing accuracy, (3) investigate heterogeneity-aware aggregation methods to handle non-i.i.d. clinical data across edge nodes, (4) implement a privacy-preserving orchestration mechanism leveraging secure enclaves and policy-driven data governance, and (5) empirically validate the framework on multi-institutional health datasets. The methodology follows a mixed-methods design combining quantitative experimentation with qualitative governance analysis. The population encompasses hospital systems, outpatient clinics, and wearable medical devices contributing time-series and imaging data relevant to disease prediction and early detection. A purposive sample of 15–20 clinical sites will participate, yielding a federated network of approximately 120 edge clients and 5 regional servers. Data collection instruments include standardized preprocessing pipelines, distributed data schemas, and task-specific labeling protocols for cardio-metabolic risk prediction, radiology anomaly detection, and sepsis warning. The study operationalizes privacy via differential privacy budgets, secure aggregation protocols, and policy-based access control. Instrument validity will be established through data lineage documentation and cross-site pilot tests. Data analysis will combine federated learning experiments with rigorous statistical evaluation. Model performance will be assessed on held-out, institution-bound test sets using AUC-ROC, F1-score, and calibration measures. Privacy leakage will be evaluated via membership inference and gradient leakage tests. The analytical framework integrates hierarchical mixed-effects models to account for site-level variability, and ablation studies to isolate contributions of privacy mechanisms, communication compression, and heterogeneity-aware aggregation. Theoretical grounding includes the fairness and accountability of AI in healthcare and the information-theoretic limits of privacy in distributed learning, with relevant reference theories such as the Information Bottleneck and Bansal and Chawla's fairness in federated settings. Anticipated findings include that edge-driven federated learning with adaptive client sampling and gradient compression achieves comparable predictive performance to centralized baselines (within 2–3 percentage points in AUC-ROC) while reducing inter-site data transfer by 60–75%. It is expected that differential privacy budgets will modestly degrade accuracy, but new aggregation strategies that mitigate client drift will recover much of the loss. The research should reveal that heterogeneity-aware aggregation yields more stable convergence in non-i.i.d. clinical data and that policy-driven governance can enhance user trust and consent compliance. The study contributes to knowledge by advancing a practical, scalable framework for privacy-preserving healthcare AI at the edge, bridging theory with deployment realities, and providing empirical benchmarks for privacy-utility trade-offs in heterogeneous healthcare settings. It highlights design patterns for secure orchestration, dynamic task assignment, and data governance that can inform stakeholders across hospitals, device manufacturers, and regulators. The main conclusion is that edge-driven federated learning can deliver clinically meaningful AI performance while enhancing privacy, governance, and latency profiles. Recommendations include adopting standardized privacy budgets and policy frameworks across institutions, integrating edge-based health analytics into existing clinical workflows, and pursuing longitudinal deployments to monitor model drift, patient outcomes, and evolving regulatory requirements. Further research should explore cross-domain federated learning with multi-modal healthcare data, advanced cryptographic techniques for secure model update propagation, and real-time explainability to support clinical decision-making.