UC Irvine

With ever-growing interests in personalized physical and mental healthcare, especially with the recent COVID-19 pandemic, along with the proliferation of applied machine learning (ML), automated health monitoring and prognosis using wearable devices with ML algorithms are increasingly more relevant. However, modern deep learning (DL) frameworks with state-of-the-art performances are unable to meet the memory and energy constraints of wearable devices. As such, there is a need for the design of efficient and effective ML models that can operate on these constrained devices while still achieving acceptable performance thresholds. This thesis presents two methodologies for effective and efficient ML under the wearable device setting, specifically a feature-augmented hybrid convolutional neural network architecture for stress monitoring using a wrist-based photoplethysmography sensor and a neural contextual-bandits-based dynamic sensor selection framework for cardiovascular disease detection with body-area-network of electrocardiogram sensor. Through extensive empirical studies, we find that our methods satisfy the constrained device settings while maintaining task performance and in some cases outperforming related ML and deep learning works. Additionally, we performed feasibility analysis on real embedded micro-controller hardware where run-time memory and energy profiling was measured and reported.