Near-Sensor Analytics and Machine Learning for Long-Term Wearable Biomedical Systems

Abstract

Wearable devices for biomedical applications have become increasingly pervasive. In a field where privacy is a major concern and latency is not well-tolerated, low-power and small-sized edge devices are of central importance. An emerging trend is to embed the processing algorithms near the sensors on the edge device to preserve privacy, reduce latency, and increase battery life. A new generation of wearable Internet of things and smart sensing systems should not only provide continuous data monitoring and acquisition but are also expected to process and make sense of the acquired data in similar ways as human experts do. Over the years, machine learning has achieved impressive results in many applications, including the biomedical field. However, the limited resources available on battery-operated devices pose enormous challenges in executing machine learning algorithms that are generally resource-demanding. In this thesis, we first evaluate and assess the ability of two representative and leading-edge ultra-low-power microcontrollers to execute machine learning models for wearable applications. We then focus on the challenging task of brain–machine interface based on the motor imagery paradigm, which allows direct communication between the human brain and external machines by merely thinking of a body part movement. We identify the state-of-the-art classification algorithms in this domain and introduce methods to reduce their computational complexity and model size allowing efficient implementation on edge devices. We further propose optimized and energy-efficient deployment techniques by exploiting hardware extensions and parallel computing. Finally, we design a new model architecture that requires significantly less memory footprint and fewer computations while at the same time keeping state-of-the-art accuracy. We additionally limit the resource requirements by proposing an effective method to reduce the dimensionality of the input data, significantly lowering the overall system’s power consumption without significantly degrading the model accuracy. With this thesis, we demonstrate for the first time that it is feasible to execute real-time inference at the edge for a brain–machine interface based on motor imagery and reach the state-of-the-art trade-off among accuracy, resource demands, and power consumption necessary for a next-generation smart wearable device.