Deep learning in resting-state electroencephalogram signals for classification of cognitively healthy older adults, Alzheimer’s disease, and Lewy body dementia patients

This study investigates the application of deep learning techniques to resting-state electroencephalogram (rsEEG) signals for the classification of cognitively healthy older adults (NOLD), patients with Alzheimer’s disease (AD), and patients with Lewy body dementia (DLB). A dataset comprising 90 participants (30 AD, 30 DLB, and 30 NOLD) was analyzed, with AD and DLB groups further divided into mild cognitive impairment (MCI) and dementia (DEM) categories. Preprocessed rsEEG signals were decomposed into standard frequency bands, and normalized power spectral density (PSD) features were extracted from 19 channels. Two deep learning models — Long Short-Term Memory (LSTM) networks and Transformer-based architectures — were implemented and validated using the Leave-One-Subject-Out (LOSO) cross-validation scheme, to simulate real clinical generalization. The LSTM classifier achieved an average accuracy of 94.97%, with 95.03% sensitivity and 94.96% specificity across three binary group comparisons (AD vs. DLB, AD vs. NOLD, and DLB vs. NOLD) and three binary category comparisons (DEM vs. MCI, DEM vs. CTRL, and MCI vs. CTRL). The Transformer model achieved even higher performance, with an average accuracy of 96.40%, sensitivity of 97.29%, and specificity of 95.56%. In multi-class classification, both approaches demonstrated robust accuracy: 86.67% (LSTM) and 87.78% (Transformer) for five-class discrimination (AD-DEM, AD-MCI, DLB-DEM, DLB-MCI, and NOLD-CTRL), and above 90% for the three-class group (AD, DLB, and NOLD) and category (DEM, MCI, and CTRL) distinctions. These results demonstrate that deep learning applied to rsEEG data can effectively distinguish between AD, DLB, MCI, and healthy aging, offering a non-invasive, accurate, and generalizable method for early detection and differential detection of dementia-related disorders.