Determination of neuromarkers associated with dementia using EEG and machine learning

Alzheimer's Disease (AD) is the most common cause of dementia and neurodegenerative disease in elderly people. Mild Cognitive Impairment (MCI) is the pre-stage of AD and is known as an early marker. Because of the severity of the disease, there is a need to develop low-cost, non-invasive, and high-resolution markers that are effective to detect the development of AD at its early stages. The target of this thesis is the determination of neuro-markers associated with dementia using EEG and machine learning. Spontaneous EEG recordings from AD/MCI groups are analyzed to identify markers of slowing, reduced complexity, and synchronization patterns within the scope of the thesis. The detection of early-stage dementia using these neuromarkers is expected to contribute significantly to the literature. In this thesis study, EEG records from Dataset 1 involving 85 HC, 85 MCI and 85 AD under eyes closed-open state, and Dataset 2 including 10 MCI and 10 HC under eyes closed state are analyzed. Permutation Entropy (PE) is proposed as complexity neuromarker, peak power amplitudes and frequencies are selected as spectral markers and wPLI between pair-wised electrodes is recommended as synchrony features. A 3-way subject based classification approach is utilized for entropy neuromarkers via. Multinomial Logictic Regression and epoch based binary classifiers are implemented to Dataset for rest of analysis. 25 of Lazypredict algorithms have considered and most relevant algorithm is picked up for further process. In this work, Dataset 2 including HC/MCI groups are also analyzed using 1D EEG segments via. state of art architectures (EEGNet, and DeepConvNet) and created 2D EEG time series are classified via. Conv2D, and ResNet architectures. Moreover, 1D EEG records are converted to 2D time-frequency representations. Then, CNN based networks and Vision Transformer are applied to classify TF images. As a conclusion, in entropy analysis, eyes open state increases MCI vs. AD classification rate in central, temporal and occipital regions within 100% acc. MCI is proven as conversion stage. Alfa and beta peak amplitudes are most promising markers for MCI vs. HC classification. Higher wPLIs are detected in HC in comparison with MCI and participants are classified up to 99% acc. All MCI and HC subjects are correctly classified within EEGNet, DeepConvNet, and Vision Transformer. Vision Transformer is superior than CNN based approaches due to its attention mechanicsm. ResNet is more effective than CNN because of using residual blocks. Current thesis study puts forward promising results for AD/MCI detection using EEG based neuro-markers and machine learning architectures.