Abstract

The worldwide population of seniors (60+) is expected to increase from 962M in 2017 to 2.1B in 2050 and 3.1B in 2100. Thus, a parallel rise in a range of age-related disorders and diseases including Parkinson’s Disease (PD) and Essential Tremor (ET) is expected. Pathological Hand Tremor (PHT) is a common symptom of such disorders, which severely affects patients’ quality of life. There are more than 100,000 Canadians living with PD today, but that number is expected to jump over the coming decade to more than 160,000, which necessitates an urgent quest for thoughtful planning and proactive measures. The development of assistive and rehabilitation technologies is one approach that stands to help affected individuals compensate for a variety of lost functionalities, and regain their self-sufficiency. One promising direction is to introduce a new communication medium to the human brain, known as Brain-Computer Interfacing (BCI), to bypass the impaired neural pathways and provide a direct link between the brain and an Assistive Device (AD). In addition, another approach is to capitalize on the remaining functionalities of the patients and compensating for the lost capabilities or correcting the flawed ones. In this dissertation, various types of signal processing and machine learning frameworks are developed to be utilized in the above-mentioned ADs, each enhancing our understanding of the problem at hand and surpassing its counterparts in terms of estimation accuracy, classification accuracy, adaptivity, and generalizability. In particular, a Bayesian optimization framework and an innovative multiclass classification scheme are developed to enhance the classification accuracy of BCI systems. In addition, two processing frameworks based on Adaptive Kalman filtering and Recurrent neural networks are introduced, which drastically enhance the accuracy of PHT estimation from the dynamics of tremorous hands. Moreover, a comprehensive screening protocol based on Convolutional Neural Networks is proposed for the diagnosis of PD from ET, which offers the state-of-the-art accuracy compared to its counterparts. Finally, a novel objective function for deep metric learning frameworks is developed to lessen the necessity of huge datasets to train artificial neural models in the biomedical domain, and in particular for neurological disorders. Such advancements would not only minimize the caregiving burden but could also help increase the number of productive and self-sufficient years patients have before debilitating disease symptoms take over.