Book Description

Recent advances in microelectronics, signal processing, Artificial Intelligence (AI), and wireless communication provide the foundation to transition from conventional in-clinic healthcare to remote health monitoring and clinical interventions. Much effort is being undertaken by the research community to design sensor-based systems such as wearables and mobile devices that provide remote monitoring capabilities for pervasive health. These embedded systems aim to enhance self-care, for example, in patients with chronic conditions such as dementia, heart failure, and diabetes. Despite the potentials, the feasibility of deploying these systems outside laboratory settings is not well established to date. An important factor limiting the large-scale adoption of remote monitoring systems is that these systems are quite resource-constrained in terms of energy capacity, memory storage, and compute power. This research aims to offer new solutions toward designing efficient embedded systems in terms of resource utilization, such as power consumption and memory usage for use in remote health monitoring. This dissertation focuses on resource efficiency from a computational perspective. It first presents a case study focused on tracking fluid intake to investigate system requirements and discover specific system properties that allow designing new techniques for improved resource efficiency. With the lessons learned from this case study, this research offers two approaches for resource-efficient system design. The first approach is a probabilistic cascading classification framework that dynamically adjusts the sampling frequency according to the learning task's complexity. In addition to adaptive sampling, the proposed framework leverages the probability of occurrence of the events to more frequently activate the machine learning module that monitors more likely events. Each machine learning module in the cascading architecture is activated by a predecessor module in the processing chain. The second approach to resource-efficient design is a classification algorithm for real-time and memory-efficient health monitoring. To this end, a hierarchical binary classification method is proposed that considers the occurrence probability of each event and constructs a decision tree that is optimized in terms of the overall depth. All the methods proposed in this research are evaluated using real data collected with wearable sensors and mobile devices.