Signal Analysis Techniques for Resource Optimization in Brain-Computer Interfaces and Other Wearables

Abstract

<p>Stroke is a serious neurological illness that often leads to motor dysfunction and affects the daily life activities of an individual. Although stroke is a complex medical issue, there is a way to reduce its impact by post-stroke neural rehabilitation. Brain-computer interfaces (BCIs) establishes a real-time interaction between the patients and rehabilitation devices and help stroke patients to restore their lost motor function. However, the existing BCIs have relatively high power consumption, more computational complexity, and large processing time, and therefore, limiting their effectiveness to enhance a patient’s health. The advances in signal processing and machine learning can help us in developing low-power and low-complexity solutions to address the aforementioned problems. Therefore, this dissertation focuses on proposing advanced signal analysis techniques for resource optimization in brain-computer interfaces and other wearables.</p> <p>The first part of this dissertation focuses on developing a patient-specific electroencephalogram (EEG) channel selection methods for motor-imagery (MI) classification using Nonnegative Matrix Factorization (NMF) which is capable of reducing computational complexity and processing time. In addition, variance activated NMF for EEG channel selection is proposed to further reduce the system complexity and effective localization of cognition during a motor task. In the MI classification framework, the theory of Riemannian geometry is utilized for feature extraction from the reduced set of EEG channels, and the neighborhood component-based feature selection algorithm is employed for feature selection. The second part of this dissertation focuses on proposing the idea of reconstruction-free compressive feature learning for wireless BCIs to leverage the resources of current BCIs by reducing the power consumption by extracting the features directly from the compressed measurements. The developed methods for low-power BCIs are validated using three MI datasets. The experimental results in this dissertation show that the developed solutions are capable of reducing the total processing time up to 95 %, the total power required up to 50 %, and the system complexity up to 75 %. Lastly, it is shown that the NMF and Compressive Sensing (CS)-based approaches developed in this dissertation are also useful for any resource constraint wearable for other low-power and real-time applications.</p>