Deep learning inspired feature engineering approach for improving EMG pattern recognition in clinical applications

Hand gesture recognition has several clinical and engineering applications in assisting people with disabilities or the aged population to enhance their standard of living. Prosthetic and assistive device applications that replace missing natural limbs with prosthetic body parts or help weak limbs are supported by gesture recognition via an artificial intelligence (AI)-enabled interface. This interfacing layer between the human (disabled person) and the machine (prosthetic) uses sensors and intelligent algorithms to interpret the electrical motor nerve signals that travel from the brain to the hand as digital orders that can be employed to control Input/Output devices. Bio-signals, especially electromyogram (EMG) (capturing muscle activities) or electroencephalogram (EEG) (capturing brain neural activities), are typically used as a source of control for these devices and interfaces. Surface EMGs are widely utilised in a control system, also called myoelectric control systems, often employing machine learning and pattern recognition (PR) systems to analyse and decipher the intentions of control from these signals. In this regard, recent literature points out the existence of several dynamic factors (e.g., varying force, forearm orientation, limb position and electrode position shift) that are potentially ignored during offline studies but clearly depicted during online clinical applications. Therefore, a substantial gap emerges in the performance of the interfacing layer between laboratory and real-time clinical applications, often attributed to the performance of the knowledge extraction algorithms.
This thesis is focused on novel methodologies that can enhance the performance of EMG pattern classification by proposing innovative, reliable and practical PR systems suitable for real-time applications, addressing robustness concerns and resolving some of the problems associated with existing PR systems. Given the substantial impact of the extracted features quality (statistical descriptors of the underlying EMG/EEG signals) on the overall PR systems, a variety of unique, resilient, hybrid and creative approaches have been established in this thesis to enhance the performance. In a broad sense, without effective feature representations, it is generally impossible to train a reliable AI model with low misclassifications; however, if the right features can be extracted, even a simple method can yield spectacular performance. This has sparked interest in feature extraction research across a wide range of academic and industrial fields. This thesis aims to develop methods for extracting features from surface electromyogram (sEMG) signals that are reliable, accurate and computationally efficient, with supporting validations of data from various clinical settings. As a result, substantive performance enhancement is demonstrated by resolving many of the problems associated with feature extraction.
The aim of this thesis is accomplished using the specific strategies or contributions outlined below.
Developing a novel recurrent fusion of time-domain descriptors method (RFTDD): This contribution focuses on the significance of information extraction in myoelectric PR from both spatial and temporal perspectives. The suggested method was compared to conventional feature extraction approaches and was found to reduce classification errors by about 12% across all subjects.
Developing a novel recurrent spatial-temporal fusion (RSTF/BiRSTF): This contribution focuses on developing algorithmic interventions to overcome the cross-sectional structure of conventional algorithms. A customised framework inspired by deep learning is presented to develop a technique that can be used in bi-directional feature extraction modes, which can be combined with any conventional feature extraction technique to enable such techniques to utilise temporal and spatial correlations on longitudinal bases. The suggested approach significantly outperforms all other methods examined compared to deep long short-term memory (LSTM) and other convolutional neural network (CNN) approaches reported in the literature. Enhancements of up to 15% were achieved on the most challenging dataset.
Developing a novel multichannel temporal cardinality feature extraction method: An innovative algorithm is proposed that uses the cardinality concept, a method for extracting low computational cost features that can be used to any number of channels. The suggested method considerably improves myoelectric PR performance, with accuracy reaching up to 99% compared to a variety of well-known approaches from the literature.
Developing a novel myoelectric PR based on the deep wavelet scattering transform (WST): This is a pioneering contribution to the possible advantages of utilising deep WST as a method to extract the feature from the EMG signal. This contribution is further supported by the application of attention mechanisms from recent deep learning frameworks that we pioneered by application with the WST in several frameworks adopted in this thesis. Comparisons are made between the results obtained using the suggested method and those obtained using other well-known feature extraction strategies. The suggested solution is 25% more accurate than the best alternative approaches. This method was also used as a practical study case to classify tremor severity, and when compared to other existing approaches in this field, the classification error rates were considerably reduced by about 12%.