An Artificial intelligence-based system for automatic diagnosis of siseases via EEG signals /

Abstract

Alzheimer’s disease (AD) is known for being the main type of dementia, distinguished by developing descent in cognition and amnesia. Early diagnosis can assist in disease management and enhance patients’ overall quality of life. Electroencephalogram (EEG) has emerged as a non-invasive tool for detecting AD that has the benefit of having a high temporal resolution. AD causes several significant changes to the patients’ EEG recordings including reduced complexity, slower EEG rhythms, and changes in synchrony. This thesis explores the use of EEG signals combined with machine learning techniques to develop a computer-aided diagnosis (CAD) tool for the improvement of AD diagnosis. Although Recurrence Quantification Analysis (RQA) features have demonstrated promising results in various EEG analysis methods such as emotion detection, they have been scarcely implemented for AD detection. In this thesis, RQA features are computed to investigate their usefulness for AD detection. Specifically, four feature groups are considered for the detection of AD from EEG recordings which are (1) RQA, (2) Hjorth, (3) Statistical, and (4) Power Spectral features. Multiple classifiers are compared including Support vector machines (SVM), K-Nearest Neighbors (KNN) and Random Forest (RF). Cross-validation (CV) methods, such as 10-fold and leave one-subject-out (LOSO) CV, are used to evaluate model performance. For investigating the relevance of features extracted from original EEG vs. from decomposed EEG, results reveal that features extracted from decomposed brain frequency sub-bands significantly enhance classification accuracy when compared to those extracted from the original EEG signal. An improvement ranging from 7% to 25% is observed for 10-fold CV and from 4% to 16% for LOSO CV. Next, two feature selection methods are applied and compared. In general, both feature selection methods yielded consistent results, leading to performance improvement by 1% to 3% in all experiments. Throughout all performed investigations, RQA features results in II the best accuracies in which its accuracies outperform Hjorth, Statistical and Power Spectral features with up to 15% and 25% for 10-fold CV and LOSO CV, respectively. These results highlight the usefulness of RQA features for AD detection from EEG signals. Best results are achieved by combining best-performing features from RQA and statistical group of features extracted from the decomposed EEG signals in which achieved accuracies were 99.2% and 96.7% for 10-fold CV and LOSO, respectively, using the SVM classifier. This research contributes to the development of more reliable AD diagnostic tools and highlights the potential of EEG-based methods in clinical practice.