Abstract

Wearable technologies are gaining considerable attention in recent years as a potential post-smartphone platform with several applications of significant engineering importance. Wearable technologies are expected to become more prevalent in a variety of areas, including modern healthcare practices, robotic prosthesis control, Artificial Reality (AR) and Virtual Reality (VR) applications, Human Machine Interface/Interaction (HMI), and remote support for patients and chronically ill patients at home. The emergence of wearable technologies can be attributed to the advancement of flexible electronic materials; the availability of advanced cloud and wireless communication systems, and; the Internet of Things (IoT) coupled with high demand from the tech-savvy population and the elderly population for healthcare management. Wearable devices in the healthcare realm gather various biological signals from the human body, among which Electrocardiogram (ECG), Photoplethysmogram (PPG), and surface Electromyogram (sEMG), are the most widely non-intrusive monitored signals. Utilizing these widely used non-intrusive signals, the primary emphasis of the proposed dissertation is on the development of advanced Machine Learning (ML), in particular Deep Learning (DL), algorithms to increase the accuracy of wearable devices in specific tasks. In this context and in the first part, using ECG and PPG bio-signals, we focus on development of accurate subject-specific solutions for continuous and cuff-less Blood Pressure (BP) monitoring. More precisely, a deep learning-based framework known as BP-Net is proposed for predicting continuous upper and lower bounds of blood pressure, respectively, known as Systolic BP (SBP) and Diastolic BP (DBP). Furthermore, by capitalizing on the fact that datasets used in recent literature are not unified and properly defined, a unified dataset is constructed from the MIMIC-I and MIMIC-III databases obtained from PhysioNet. In the second part, we focus on hand gesture recognition utilizing sEMG signals, which have the potential to be used in the myoelectric prostheses control systems or decoding Myo Armbands data to interpret human intent in AR/VR environments. Capitalizing on the recent advances in hybrid architectures and Transformers in different applications, we aim to enhance the accuracy of sEMG-based hand gesture recognition by introducing a hybrid architecture based on Transformers, referred to as the Transformer for Hand Gesture Recognition (TraHGR). In particular, the TraHGR architecture consists of two parallel paths followed by a linear layer that acts as a fusion center to integrate the advantage of each module. The ultimate goal of this work is to increase the accuracy of gesture classifications, which could be a major step towards the development of more advanced HMI systems that can improve the quality of life for people with disabilities or enhance the user experience in AR/VR applications. Besides improving accuracy, decreasing the number of parameters in the Deep Neural Network (DNN) architectures plays an important role in wearable devices. In other words, to achieve the highest possible accuracy, complicated and heavy-weighted Deep Neural Networks (DNNs) are typically developed, which restricts their practical application in low-power and resource-constrained wearable systems. Therefore, in our next attempt, we propose a lightweight hybrid architecture based on the Convolutional Neural Network (CNN) and attention mechanism, referred to as Hierarchical Depth-wise Convolution along with the Attention Mechanism (HDCAM), to effectively extract local and global representations of the input. The key objective behind the design of HDCAM was to ensure its resource efficiency while maintaining comparable or better performance than the current state-of-the-art methods.