UC Irvine

With the advancement of technology and the prevalence of Internet of Things (IoT), wearable devices have been gaining huge momentum as consumer devices over the past few years. The wide adoption and multiple sensor modalities in those wearable devices have enabled them to be used for designing different digital health solutions for continuous and remote monitoring. The continuous use of wearable devices results in a myriad volume of physiological data being collected from multiple sensors. Transferring this collected data from wearable devices (edge) to server (cloud) introduces three major challenges - increased energy consumption, increased latency, and vulnerability to breach of privacy. To tackle these challenges, on-device computing solutions have been adopted for different wearable healthcare systems. However, the small form factor of wearable devices imposes three constraints on the on-device solutions. The solutions should be energy-efficient, memory-efficient, and provide maximum performance within the previous two constraints. Moreover, getting the ground truth label of the collected data from the user to update the classification models is a challenging task without user involvement. Furthermore, fusing data from multiple heterogeneous sensors may often introduce new challenges in terms of achieving maximum performance regardless of computing architecture.

To address these challenges, this thesis presents some efficient methodologies for different wearable healthcare applications like - Myocardial Infarction Detection, Human Activity Recognition, Stress Detection, and Eating Activity Recognition. The first part of the thesis explores the Myocardial Infarction Detection application where we propose a Template Matching based Early Exit architecture that achieves energy and memory efficiency while outperforming state-of-the-art work. The second part proposes an adaptive convolutional neural network for Human Activity Recognition application. Our proposed solution outperforms state-of-the-art works while achieving energy and memory efficiency. In the third part of the thesis, we explore the Eating Activity Recognition application where we propose an online learning methodology to adapt to changing eating habits of the user. Our proposed online learned neural network outperforms other competitive offline methods while being energy-efficient. The fourth and final part of the thesis explores Stress Detection application using multiple heterogeneous sensor modalities. We propose a selective fusion approach using context-aware sensor selection from wearable devices which achieves better performance compared to state-of-the-art works.