Abstract Alzheimer’s disease (AD) is a progressive form of dementia. Electroencephalography (EEG) offers a promising avenue for AD diagnosis and differentiation from other dementias, but EEG data complexity and noise have posed challenges. This study proposes FPSD-CNN, integrating Frequency Power Spectral Density (PSD) analysis with a deep convolutional neural network (CNN) to enhance AD classification accuracy from EEG signals. We perform binary classification on a public EEG dataset: AD vs. cognitively normal (CN) subjects, and CN vs. frontotemporal dementia (FTD) subjects. The dataset includes 88 subjects, with 36 AD patients (mean age=66.4, sd=7.9), 23 FTD patients (mean age=63.6, sd=8.2), and 29 CN subjects (mean age=67.9, sd=5.4). PSD features are extracted via frequency-domain analysis, flattened into FPSD features, and input to a CNN for automated extraction and classification. The model is trained and tested on raw EEG and artifact-removed data to assess noise robustness. On raw EEG for AD vs. CN, FPSD-CNN achieves 88.32% accuracy; for CN vs. FTD, it reaches 82.22%. This method demonstrates superior accuracy and noise resilience, positioning it as a promising non-invasive approach for early detection and monitoring of Alzheimer’s disease. Code and models are available at GitHub: changjingzhi/FPSD-CNN. Graphic Abstract
Li et al. (Wed,) studied this question.