Privacy-preserving heterogeneous multi-modal sensor data fusion via federated learning for smart healthcare

Jing Wang, Mohammad Tabrez Quasim, Bo Yi

Abstract: Highlights•We propose a novel privacy-preserving framework for heterogeneous multi-modal sensor fusion based on federated learning, which enables healthcare institutions to collaboratively train shared diagnostic models without exchanging raw sensor data while effectively handling diverse medical sensor modalities, including physiological signals, environmental parameters, and biochemical measurements.•We develop an adaptive tensor-based fusion mechanism that automatically handles varying combinations of sensor modalities across healthcare institutions. Due to the use of innovative tensor decomposition techniques, our approach captures the intricate relationships among different sensor types while still ensuring an adequate level of privacy, thus allowing for proper patient supervision without losing sensitive information.•We introduce a privacy-preserving federated learning architecture designed for heterogeneous medical sensor networks. Our framework has built-in secure parameter aggregation techniques, simulating knowledge sharing across multiple healthcare facilities without jeopardizing data privacy.