CrypTen-FL: A Secure Federated Learning Framework for Multi-Disease Prediction from MIMIC-IV Using Encrypted EHRs

Abstract

The increasing demand for privacy-preserving machine learning in healthcare has driven the need for federated approaches that ensure data confidentiality across institutions. In this work, we present CrypTen-FL, a secure federated learning framework for disease prediction using the MIMIC-IV electronic health record (EHR) dataset. CrypTen-FL enables collaborative model training across multiple hospitals without sharing raw patient data, thereby addressing critical privacy concerns through the integration of Secure Multi-Party Computation (SMPC) using CrypTen and differential privacy mechanisms. We adopt a Transformer-based neural architecture to effectively capture the temporal and high-dimensional nature of EHR data, enabling accurate prediction of multiple clinically significant conditions. The framework incorporates decentralized key generation, secure aggregation, and cross-institutional evaluation to assess generalization performance and robustness. Experimental results demonstrate that CrypTen-FL achieves competitive predictive performance while offering strong privacy guarantees, paving the way for secure and scalable AI applications in real-world healthcare settings.