COMPUTATIONAL ALGORITHMS FOR IMPROVING ENERGY EFFICIENCY AND RELIABILITY IN WEARABLES

Ramin Fallahzadeh

Recent advances in electronics, machine learning and signal processing, and wireless communication have laid the groundwork for moving from traditional in-clinic healthcare to remote settings. Recently, the use of Internet-of-Things (IoT) solutions for intelligent and pervasive healthcare has gained growing popularity in research community. Various remote health monitoring systems have been developed by researchers for improving self-care in patients with chronic medical conditions such as heart-failure and diabetes. These efforts, however, have not been as successful in gaining commercialized popularity when used outside the lab settings. Due to lack of user adherence to such monitoring and interventional technologies, the successful large-scale deployment of these solutions has remained a major challenge to date. Wearable technologies, as a large part of IoT, have attracted considerable attention mainly due to their increasing capacity for continuous and high precision symptom and behavioral monitoring. This work, specifically, focuses on wearable monitoring technologies and physical activity monitoring as its core application. A large body of research has been conducted to identify the major contributing factors to low user-compliance in wearable technologies. Inspired by the prior findings, in this work, we propose several solutions that can contribute to improving technology compliance by emphasizing energy efficiency and computational reliability. Two key factors have been considered: (1) power-aware design; and (2) computational autonomy. We propose a robust and low-power adaptive compressed sensing framework that significantly improves battery lifetime of wearable sensing nodes. Our framework optimizes the energy consumption of the wearable node by adaptively adjusting the sampling rate according to the context e.g., signal-types and on-body sensor location. The proposed framework is evaluated on real-world data and on an actual wearable node for accurate measurement of performance and energy savings. In addition, we overview the concept of computational autonomy and its application in wearable activity recognition. We propose a reliable cross-user transfer learning approach that leverages the high-level knowledge extraction and mapping across subjects’ sensor data to boost the base supervised model. Our experimental results on multiple real-world datasets demonstrate the effectiveness of this novel cross-subject transfer learning approach in improving accuracy of the activity recognition model.

COMPUTATIONAL ALGORITHMS FOR IMPROVING ENERGY EFFICIENCY AND RELIABILITY IN WEARABLES

Files and links (1)

Abstract

Metrics

Details