Automated electromyogram signal classification frameworks based on singular spectrum analysis variants

Abstract

The early diagnosis of neuromuscular disorders and the development of assistive technologies rely heavily on the accurate analysis of electromyogram (EMG) signals, which are inherently non-stationary, noisy, and complex. This thesis introduces novel frameworks that integrate advanced singular spectrum analysis (SSA) variants—namely, automatic SSA (Auto-SSA), sliding mode SSA (SM-SSA), and multivariate SM-SSA (MSSA)—with quantum convolutional neural networks (QCNNs) to enhance classification performance and robustness in biomedical signal processing tasks. The proposed frameworks are validated across three key applications. For amyotrophic lateral sclerosis (ALS) detection, intramuscular EMG signals from benchmark datasets were decomposed using Auto-SSA, an adaptive method that selects decomposition parameters automatically, extracting most significant features of the signal via particle swarm optimization (PSO). The resultant QCNN classifier achieved a testing accuracy of 98.50% on 200 training samples, outperforming conventional classifiers. For eye movement detection, extraocular muscle EMG signals were analyzed using SM-SSA, which segments signals into overlapping frames for automated decomposition, followed by neighborhood component analysis (NCA)-based feature selection. The QCNN framework yielded a remarkable 98.70% accuracy on a publicly available six-class eye movement dataset with 256 training samples. For human activity recognition (HAR), multichannel surface EMG signals were processed using MSSA, which preserves inter-channel dependencies through channel-aligned decomposition. Multi-domain features—spanning time, frequency, and entropy characteristics—were extracted and refined using minimum redundancy maximum relevance (MRMR) before classification with a 10-qubit QCNN. The HAR framework achieved superior classification accuracies of 98.81%, 98.78%, and 98.86% for aggressive, normal, and combined activity classes, respectively, using an extensive dataset comprising 20 physical actions