Design and Development of Signal Processing and AI-Based Classification Framework for BCI Applications

Abstract

Brain Computer Interface (BCI) systems represent a revolutionary advancement in neural output manipulation, empowering individuals to interact with their environment solely through brain signals, bypassing conventional neuromuscular pathways. The systems are designed primarily to control assistive devices such as speech synthesisers and robotic wheelchairs, facilitating seamless interaction with the external surroundings. Preprocessing and feature extraction modules form the core of BCI systems, which play pivotal roles in enhancing system performance. To this end, this work endeavours to pioneer novel methodologies to extract and classify EEG signals, aiming to enhance the efficacy of BCI systems. Beginning with the preprocessing phase, the focus is on optimising key parameters, including time segment, frequency band determination, and the spatial arrangement of electrodes, to design an efficient model for MI- EEG signal classification. This method strategically utilises data from a minimal set of EEG channels, optimising computational efficiency without compromising performance. An automated approach incorporating the time-reassigned multisynchrosqueezing transform technique with the deep learning framework, E-CNNet, has been developed to extract and classify distinctive MI-EEG features. The time-reassigned multisynchrosqueezing transform technique extracts time and frequency information from EEG signals, transforming MI-EEG signals into time-frequency representations with enhanced resolution. The developed E- CNNet model captures meaningful patterns from the time-frequency representations for precise feature extraction, subsequently classified by an ensemble of classifiers. This approach achieved high classification accuracy and enhanced the overall reliability of the BCI systems. The reduction in the methodological complexity improves the viability of the proposed methodology for practical-world BCI applications. Next, an innovative method tailored for feature extraction and the classification of MI-EEG signals for BCI systems is introduced. The conjunction of the scaling-basis chirplet transform along with the hybrid parallel-series attention-driven deep learning architecture presents substantial improvement in MI-EEG signal analysis. The scaling-basis chirplet transform effectively maps non-stationary MI-EEG signals to high-resolution and energy-concentrated time-frequency representation, thereby enhancing the efficacy of the model by capturing intricate dynamics of neural activities. The hybrid parallel-series attention-driven model is designed for extraction of multi-scale information imperative to classify MI-EEG signals efficiently. This framework enhances the extraction of discriminative features and facilitates more accurate classification of neural patterns associated with MI tasks. This method attains superior classification accuracy, underscoring its viability for real-time BCI applications, making it a viable solution for practical, real-time BCI applications. With further advancement, the graph theory has been leveraged to construct brain functional networks, facilitating the identification of MI-EEG signals by modeling the complex neural interactions and connectivity patterns among distinct brain regions. To mitigate noise and minimize redundant information present in the EEG channels, a novel method, Stockwell Transform- based phase lag index-Wilcoxon signed test (S-PLI-WT) is proposed. The PLIiv computed via S-transform enhances time-frequency resolution, capturing inter-channel MI- EEG information by computing EEG phase across frequencies and time points. This method uses phase lag combined with a weighing mechanism which eliminates volume conduction and noise, identifying significant connections without thresholds via the Wilcoxon signed- rank test method. The global and local network topology attributes are extracted from selected EEG channels. The study further develops the RFE-ELI5 algorithm to select significant features. It employs the Explain Like I’m 5 (ELI-5) framework of Explainable Artificial Intelligence (XAI) to interpret the outcomes of the proposed framework and offers insights into the features that play a significant role in the decision-making process of the proposed framework. Furthermore, the study analyses quantum machine learning to classify MI-EEG signals (left-hand, right-hand, feet and tongue MI classes). Furthermore, the study delves into the deployment of quantum machine learning algorithms to classify MI-EEG signals, thereby leveraging the inherent advantages of quantum computing in processing complex EEG datasets. The proposed methodology outperformed existing state-of-the-art models developed for classifying motor imagery data, thereby affirming its efficacy in advancing MI-EEG signal analysis for BCI systems. This extensive exploration delves into the forefront of BCI systems, revolutionising the manipulation of neural outputs and significantly enhancing the user interaction with the environment. By emphasising innovative preprocessing, feature extraction and classification methodologies, we have markedly enhanced the capabilities of BCI systems, facilitating more accurate interpretation of neural signals and enhancing overall system functionality. Our proposed channel selection method, leveraging brain functional connectivity, ensures robust neural signal processing, preserving vital neurological activity for comprehensive analysis. Furthermore, the techniques developed to extract and classify features from MI-EEG signals showcase remarkable accuracy and efficiency, leveraging minimal EEG channels while capturing complex temporal and spectral dynamics. These advancements underscore the potential for real-time BCI applications, paving the way for more feasible and effective BCI system designs with profound implications for neuroscientific research and clinical practice.