Abstract
The digitization of healthcare has generated massive repositories of Electronic Health Records (EHRs), offering unprecedented opportunities for deep learning models to improve predictive clinical analytics. However, centralized aggregation of sensitive clinical records poses severe patient privacy risks and regulatory challenges under HIPAA and GDPR. Federated learning (FL) enables collaborative model training across decentralized health institutions without transferring raw patient data, yet recent attacks demonstrate that FL models remain susceptible to membership inference and gradient inversion. To resolve this critical tension between data privacy and analytical utility, this study proposes a privacy-preserving federated Transformer framework incorporating Differential Privacy (DP). We integrate Differentially Private Stochastic Gradient Descent (DP-SGD) with an adaptive per-sample gradient clipping mechanism adapted for parameter-heavy Transformer architectures trained on clinical text and structured EHR sequences. Evaluating our framework across ten simulated hospital nodes utilizing the MIMIC-IV and eICU benchmarks for mortality and 30-day readmission prediction, we demonstrate that high clinical utility can be preserved under rigorous privacy guarantees. Specifically, our model achieves an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.884 at a formal privacy budget of $\epsilon = 2.0$ and $\delta = 10^{-5}$, outperforming conventional private federated baselines by 4.2%. These results establish a scalable, mathematically rigorous pathway for multi-institutional health data analytics without compromising individual patient confidentiality.
