8th ISF

COMP035 - Machine Learning Framework for Dementia Classification using Electroencephalogram (EEG)

Dementia, affecting approximately 55 million people worldwide, is a progressive condition characterized by a decline in cognitive function, significantly impacting individuals and society. Although there is no cure, early detection is critical for providing effective treatment and improving quality of life. Electroencephalogram (EEG) measurements, which capture the brain's electrical activity, have shown potential as a non-invasive method for diagnosing dementia and identifying its stages. This research utilizes machine learning techniques to analyze EEG signals and classify subjects into three groups: normal condition (NC), mild cognitive impairment (MCI), and dementia (DEM). The data, collected during visually stimulated tasks (VST) at Vidyasirimedhi Institute and Siriraj Hospital, offers insights into cognitive processes involving perception and decision-making. These tasks simulate real-world cognitive challenges, enabling a better understanding of dementia-related impairments. To enhance cost-efficiency, the framework reduces the number of channel electrodes used. Based on a literature review, six channels—C3, C4, Cz, CP3, CP4, and CPz—were selected as the most relevant for detecting dementia-related changes in brain activity. Additionally, the analysis focuses solely on aperiodic components derived from power spectral density (PSD), using slope and intercept features for classification, minimizing data collection efforts while maintaining accuracy. Multiple algorithms were employed, including traditional methods such as support vector machine (SVM), k-nearest neighbors (KNN), Naive Bayes, and random forest, as well as neural network-based models like EEGNet and ShallowConvNet. Patient-based classification, accounting for individual variability in EEG signals, was adopted to ensure applicability in real-world scenarios. By addressing subject-specific variations, the framework improves the robustness of dementia detection. The best results were obtained using Naive Bayes with aperiodic PSD features, achieving a maximum accuracy of 0.700 and sensitivity of 0.760. While neural network-based methods demonstrated potential, traditional algorithms provided competitive performance with lower computational complexity, making them suitable for resource-constrained environments. Moreover, the reduced channel selection offers practical advantages by simplifying EEG data acquisition while maintaining diagnostic efficiency. By optimizing channel selection and leveraging aperiodic signal features, the framework demonstrates the potential to balance diagnostic accuracy with practical implementation. However, further refinement, validation, and larger-scale testing are essential before clinical deployment.

Show More