Privacy-Preserving Framework for Multi-Institutional Medical Time-Series Analysis via Homomorphic Encryption: Design...

Area of Science:

• Medical Artificial Intelligence
• Data Privacy and Security
• Longitudinal Data Analysis

Background:

• Medical AI development requires large, multi-institutional datasets, posing privacy risks and hindering data aggregation.
• Navigating the utility-privacy trade-off is challenging due to sensitive information leakage concerns.

Purpose of the Study:

• To design a secure multiparty deep learning system for privacy-preserving modeling of distributed medical time-series data.
• To achieve predictive accuracy comparable to non-secure models while ensuring strong security and efficiency.

Main Methods:

• Developed a framework using threshold homomorphic encryption for secure recurrent neural network training on distributed longitudinal data.
• Implemented an optimized encrypted matrix multiplication scheme, a secure ciphertext refresh protocol, and lightweight encryption parameters.
• Evaluated the system on four intensive care unit datasets for mortality and sepsis prediction.

Main Results:

• The system demonstrated practical efficiency, with ~1 minute per training iteration for 125 local batches.
• Securely trained models achieved comparable or superior predictive performance to non-secure centralized models.
• Achieved an AUC of 0.8480 on the PhysioNet Challenge 2012 dataset, outperforming the non-secure baseline (0.8404).

Conclusions:

• Presents a viable and efficient solution for cross-institutional, privacy-preserving analysis of longitudinal medical data.
• Successfully bridges the utility-privacy gap, enabling safer collaborative research.
• Facilitates robust knowledge discovery and decision support while adhering to data protection standards.