Towards Diagnosis of Dementia: Microstate Analysis of EEG Signals and Classification using Machine Learning

Abstract

Dementia is a neurodegenerative disorder characterized by cognitive decline which presents significant challenges to healthcare systems around the world. Early and accurate detection is important for effective intervention and management. Electroencephalography (EEG) provides a noninvasive and cost-effective technique for assessing neurological conditions. This thesis explores the use of EEG microstates as features for traditional machine learning to detect dementia.

The study develops two machine learning techniques: one employing traditional machine learning for microstate features and the other training a deep learning model using topoplot images based on GFP peaks. EEG microstates are quasi-stable states representing transient functional brain activities that provide valuable insights into brain activity associated with dementia. By leveraging both the temporal and spatial resolution of EEG microstates, the analysis reveals patterns linked to dementia. Traditional machine learning models such as Support Vector Machines (SVM), Random Forests (RF), and eXtreme Gradient Boosting (XGB) were deployed to classify Dementia, Mild Cognitive Impairment (MCI), and healthy individuals (Normal).

In addition, we proposed extracting topographic images from EEG signals and utilizing a deep learning model to classify EEG recordings. The proposed five-layer 2D Convolutional Neural Network (CNN) is a simple, lightweight model. The CNN kernel captures spatial features from images, enhancing the model's ability to differentiate between dementia-related brain dysfunction and healthy brain activity. The model aims to classify EEG signals into three categories: Dementia, MCI, and healthy individuals.

Furthermore, we compared our model with baseline results from previous work conducted on the publicly available CAUEEG dataset. Our findings demonstrate the effectiveness of the traditional model using microstate features, achieving 74 % accuracy for classifying Dementia, MCI, and Normal, and 76% for binary classification. The proposed deep learning model achieved 82% accuracy for multiclass classification. Overall, this thesis contributes to dementia diagnosis by leveraging EEG microstate analysis and advanced machine learning techniques.