Drift-Aligned Personalization for Privacy-Preserving Edge Health Monitoring

Abstract

Health surveillance systems that are based on edges make it viable to continually monitor physiological data, minimize latency, and minimize the use of bandwidth. Nonetheless, the concept drift is also sensitive to edge-device personalized prediction due to variations in the patient conditions, sensor variations and changes in environmental conditions. In the meantime, another requirement of a distributed healthcare ecosystem is patient privacy. The paper also proposes a privacy preserving Drift-Aligned Personalization Framework (DAPF) to privacy conscious edge health monitoring derived based on adaptive learning, drift aware personalization and secure collaborative intelligence. It is a compilation of six large-scale layers Wearable Data Acquisition, Edge Preprocessing, Concept Drift Detection, Personalized incremental learning, Privacy Preserving Federated Aggregation and Clinical Decision Support. Streams of real-time physiologic data are constantly monitored to identify distribution changes and then lightweight personalized models are updated on-the-fly on a device without being sent sensitive patient data. Secure federated learning applies the parameteric encryption of model parameters to be aggregated periodically to improve global knowledge and keep local personalization. Differential privacy and secure aggregation mechanisms guarantee the confidentiality and adherence to the regulations. The proposed framework brings accurate predictions, less disastrous catastrophic degradation of the model due to drift, low overhead of communication and the heterogeneous edge devices can be scaled up. Essentially dedicated adaptive intelligence + privacy-conscious distributed learning: DAPF can provide an established foundation of intelligence based healthcare systems over the next generation which can provide high quality, flexible and trusted care of remote patients.