Wearable medical devices can continuously monitor individuals' health to enable personalized, preventative medicine. However, before such devices can inform medical decisions, they face multifaceted and often contradictory design objectives. Specifically, they must provide accurate medical information that is trusted by doctors and patients. At the same time, they must be portable, unobtrusive, have a long battery lifetime, and preserve users' privacy. In this thesis, I propose a wearable medical device development strategy based on software-hardware co-design that leverages Edge Artificial Intelligence (edge AI) to enhance data privacy.
To showcase this co-design methodology, I first develop three novel algorithms to address unmet medical needs. Each one is designed to be reproducible and explainable to foster trust between the Machine Learning (ML) and medical communities. The first application aims to diagnose respiratory conditions, such as COVID-19, using ML classification of cough audio samples. Moreover, I develop a novel methodology to improve the consistency of crowdsourced datasets, resulting in a 15.6% increase in model accuracy. Second, I contribute a ML model based on audio signals to quantify the daily coughs of patients with chronic cough disorders. This approach achieved a cough event detection F-1 score of 82%. Finally, I design a model that exploits multimodal biosignals to monitor changes in subjects' emotional states, culminating in a classifier with an F-1 score of 79%.
The next step in my co-design approach is to optimize these algorithms for energy-efficient edge AI execution on embedded systems. To this end, I propose two novel optimization strategies that leverage multimodality and runtime adaptability to strike an optimal balance between accuracy and efficiency. First, I develop the Cough-E algorithm, which changes between low-energy and high-precision sensing modalities to perform real-time cough detection. This method reduced the energy consumption of the algorithm by over 70% for only a 1.3% decrease in F-1 score. Next, I contribute the ACE methodology, which automatically transforms a ML classifier into models of increasing complexity, selecting the appropriate one at runtime. The ACE method enabled a 49% runtime speedup of the emotional state classifier, for only a 0.5% F-1 score loss.
Next, the hardware of the system is advanced using novel programmable computing platforms that can efficiently accelerate computing bottlenecks of the applications. To this end, I develop the DISCO-CGRA, an innovative design that accelerates key stages of the biomedical edge AI applications. I perform a physical design space exploration of the DISCO-CGRA and compare it to state-of-the-art edge AI computing platforms in deeply scaled transistor technologies. As a result, the DISCO-CGRA uses 22.5x less energy to compute common biomedical functions than state-of-the-art platforms. Additionally, implementation in future technology no
