Optimized algorithms and machine learning techniques for biosignal processing on ultra-low-power computing platforms

In recent years, advancements in electronic systems have driven the development of implantable and wearable devices that facilitate continuous health monitoring through the extraction of physiological parameters from biosignals, such as Electrocardiogram (ECG) and Electroencephalogram (EEG). Biosignal-based applications have become central to various fields, ranging from fitness to medical-grade diagnostics. However, implementing biosignal processing presents significant challenges, notably in achieving a balance between computational power and energy efficiency, which is essential for extended battery life in portable devices. This thesis contributes to this field by presenting a framework of end-to-end methodologies designed to optimize energy efficiency in executing computationally intensive signal processing tasks on resource-constrained embedded devices. Through a combination of optimized system architectures, low-power processing strategies, and machine learning-based algorithms, the thesis offers novel solutions for achieving high-performance ExG signal analysis within strict energy budgets. Key aspects include the design of Analog Front Ends (AFEs) to ensure high-fidelity signal capture with minimal energy draw, as well as optimizing digital processors to handle complex operations such as filtering, feature extraction, and pattern classification within limited memory and processing power. Additionally, this research explores the adaptation of machine learning algorithms, such as CNNs and TCNs, for edge-based biosignal processing, emphasizing model compression to reduce computational overhead. The research demonstrates a sustainable solution for real-time biosignal processing on ultra-low-power (ULP) parallel platforms, offering significant advantages over traditional MCUs in both energy efficiency and processing capability. To validate the proposed methodologies, the thesis investigates two primary case studies. The first focuses on ECG signal processing and classification, showcasing how on-device computation minimizes data transmission and latency, thereby improving privacy, energy efficiency, and responsiveness. The second evaluates ear-EEG as a promising alternative to conventional, full-scalp EEG, demonstrating its viability in mobile health applications.