Tawhid, Nurul Ahad (2023) Automatic Detection of Neurological Disorders using Brain Signal Data. PhD thesis, Victoria University.

Abstract

Managing neurological disorders is a major challenge for public health and health care systems in Australia and around the world. Currently, there is no reliable way of identifying disorders from brain signal data automatically, quickly, and accurately. Electroencephalography (EEG) is a powerful and popular technique to capture brain signal data for neurological disorder diagnosis through visual inspection. But this process is time-consuming, subjective, exhaustive, and error-prone. EEG records the electrical activities of the brain and provides important information about changes in electrophysiological brain dynamics for neurological diseases including autism spectrum disorder (ASD), schizophrenia (SZ), epilepsy, and Alzheimer’s disease. While EEG signals provide substantial insight into brain activity, there is a limited body of research dedicated to the automated detection and assessment of various neurological diseases and disorders. Even today, experts frequently evaluate the EEG signal manually. Therefore, it is necessary to develop a computer-aided diagnostic (CAD) system for the precise and automatic diagnosis of neurological disorders as early as possible. Classification methods play a crucial role in distinguishing EEG segments and assessing an individual’s health status. The effective utilisation of appropriate classification algorithms to accurately and efficiently identify distinct EEG signals associated with various disorders poses a significant challenge in designing a reliable and efficient CAD system. This study intends to work towards the detection of two neurological disorders, named ASD and SZ. Existing research works related to these two diseases have some limitations, such as: Research problem 1: Those are still insufficient and have scope to improve in terms of accuracy and performance. Research problem 2: Very few studies have considered developing a system for classification of multiple neurological disorders in a single framework. Research problem 3: Most of the studies are related to a particular disease and verified using a specific dataset, which left questions about their effectiveness on other datasets of the same disease as well as their efficacy on other diseases. Research problem 4: Lack of CAD systems for assisting clinicians in diagnosis of those diseases. The key aim of this project is to address the issues mentioned above by developing several innovative frameworks that will use EEG data to automatically and efficiently identify ASD and SZ. We have used several publicly available datasets for validation of the proposed methods. To address research problem 1, we have proposed a time-frequency spectrogram imagebased framework for ASD classification using both machine learning (ML) and deep learning (DL)-based classification techniques (Chapter 3). In this technique, EEG signals are first converted into spectrogram images using the short-time Fourier transform (STFT), and then those images are used as input for different ML and DL-based classifiers. Experimental results show that both ML and DL methods performed well in the EEG signal classification between ASD and healthy control (HC) subjects, but DL performed better than the ML-based classifiers. The research finding also indicates that the proposed method can be used for developing a CAD system for the identification of ASD from HC subjects. Similarly, for SZ detection, we have developed a framework using an entropy topographic image with a DL-based convolutional neural network (CNN) model (Chapter 4). We used Shannon entropy to extract entropy values from each channel of the EEG signal and then plotted them on the brain scalp to produce the topographic image. Then those images are trained and classified using our proposed CNN model. The obtained results indicate that the proposed method can be used for brain signal data mining purposes. The second research problem motivates us to propose our third research work: developing a multi-class classifier for multiple neurological disorders using spectrogram images of EEG signal data (Chapter 5). In this technique, we have extended our proposed method of ASD classification to a five-class classification framework. In this method, we have classified four neurological disorders, namely ASD, SZ, epilepsy, and Parkinson’s disease, from HC subjects. We have used two histogram-based feature extractors and four ML-based classifiers to categorise those extracted features. We have also used DL-based models for the classification of those images, and the obtained result shows a promising outcome. To solve the third research problem, we have developed a generic CNN model for the classification of EEG data for different neurological disorders (Chapter 6). Most of the previous frameworks worked for a particular disease and a particular EEG dataset, which motivated us to create a generic CNN model that can work with different neurological disease classifications. The experimental results show promising outcomes on different datasets from various neurological diseases. Finally, to resolve the first and fourth research problem, we have proposed a framework for subject-independent SZ detection using a DL-based Convolutional Long Short-Term Memory (ConvLSTM) model (Chapter 7). The proposed model is designed to perform classification independently of training and testing subjects so that it can perform classification in real-life situations where new testing EEG data is unknown to the trained model. Finally, we have developed a web-based CAD system for SZ detection using the proposed ConvLSTM-based framework. As of now, the outcome of this PhD work is four journal articles and two conference papers related to the proposed methods that have been published in reputed journals and conferences. One more article is currently under review. Also, the developed webplatform- based CAD system will be a helpful tool for the clinical diagnosis of different neurological disorders.

Search Google Scholar

Repository staff login