The human upper limb is a complicated system with many degrees of freedom. Its dexterity enables human to complete various complex motor tasks in a precise and efficient way. Understanding how the central nervous system (CNS) generates the movement and coordinates the muscles could be beneficial in both skill learning and robotic design. Considering the safety issue, the methods to measure human brain and muscle activities are normally limited to be non-invasive. Electroencephalography (EEG) and electromyography (EMG) are two of the most commonly utilized methods which were also used for researches in this thesis.

EEG decoding has long been a popular research topic. Though deep learning methods, such as the convolutional neural network (CNN), could provide a high accuracy in application, they were not always the best choice due to the low explainability and limited insight into the motor control and neuromuscular coordination. Thus, in the first part of our research, we attempted to decode the movement status (moving or static) of subjects from their EEG recordings during a grasp-and-lift task using common spatial patterns (CSP) for feature extraction and linear discriminant analysis (LDA) for classification. At the same time, we further generated several surrogate data sets to evaluate the model. According to our results, the real EEG data had smaller variance but similar classification accuracy compared with the surrogate data generated according to preprocessed EEG using cross-validation. Our results revealed a possibility that though the model achieved an accuracy higher than 50% in binary classification, it could still only capture limited information from the EEG data. Introducing the extra tests to evaluate the models is helpful especially when it involves a more complex dynamic process.

According to our results about the detection of movement status, decoding movement from EEG directly is challenging. Thus, it may be helpful to introduce an intermediate layer. Considering the fact that the CNS actually interacts with the muscles to control the body, the muscles could become the “bridge” linking the CNS and the generated movement. Thus, we moved on to infer the motor control strategy from surface EMG, which is one of the most commonly utilized non-invasive techniques to measure muscle activities. It is distinct and clear compared with EEG and could provide a straightforward view of motor control. To better understand the motor control strategies, we conducted experiments employing piano playing, a complicated, highly-skilled motor task as the experimental task. We recorded the EMG signals from 14 muscles on the left and right side respectively during the subjects' performance and factorized the motor modules, known as muscle synergies, using non-negative matrix factorization (NMF) algorithm. To study the variation of motor control under different constraints and requirements, we required our subjects to play designated scales and pieces with the combination of two different levels of tempo and contact force; to look into the alteration brought by expertise, we invited both expert pianists and novices to participate in the experiment. Our results revealed that both groups generally utilized similar muscle synergies but showed different preference in them. When adapting to different tempo and contact force requirements, the experts had more stable synergy selection compared with the novices. Apart from the alteration in muscle synergy selection, which is more general, we also identified the fine modulation in the balance of the components within the muscle synergies when the subjects were required to switch between specific styles. Compared with the novices, experts were found to have a much larger repertoire in fine-tuning the weights of muscle components in tempo switching.

Finally, we dug further into the motor modules decoded from EMG. Considering the intuition that the muscles could be either relaxed or contracted, the weights of muscles in the synergies are normally regarded to be non-negative. The most commonly utilized algorithm for synergy extraction, NMF, also has a non-negative constraint on the extracted muscle synergies. However, as related researches progressed, the inhibition in motor control has been proposed by several studies, which may introduce the components with negative weights into the muscle synergies. These components were referred to as inhibitory components, which were missed in synergy extraction using NMF. We then utilized a rectified latent variable model (RLVM) without the nonnegative constraint on muscle synergies. We also utilized NMF synergies as initialization to make full use of the prior knowledge. Our results revealed a higher proportion of inhibitory components in older subjects and more senior experts. In the inhibitory components commonly modulated, less negative weights were found with increased tempo or decreased force.

In summary, our research reveals the limitation of the traditional feature in EEG decoding, and emphasizes the importance of model evaluation especially when dealing with EEG data recorded during more complex movements. In EMG analysis, our research shed light on dynamic motor strategies in complex, highly-skilled motor tasks by identifying the alteration in motor modules related to expertise and adaptation to different constraints and requirements. Furthermore, the detection and analysis on the potential inhibitory components which were long neglected by previous researches brought new insights about human motor control. In new skill learning, our findings about the expertise-related alteration in muscle synergies could help to provide applicable suggestions to improve the practice efficiency. Meanwhile, the human upper limb, which is one of the most flexible and complicated systems in the world, could serve as a reference in motor prosthetic design and robotic control. The stable recruitment of modules, fine-tuning of internal components and the introduction of inhibitory components identified in our work could help enhance the performance and functionality.