Muti-band Morlet mutual information functional connectivity for classifying Alzheimer's disease and frontotemporal dementia with a deep learning technique

Accurate differentiation of Alzheimer's disease (AD), frontotemporal dementia (FTD), and healthy control (HC) is critical for early diagnosis and intervention of brain disorders. This study introduces a deep learning framework that leverages electroencephalography (EEG)-derived multiband functional connectivity (FC) features. Multiband Morlet wavelet mutual information (MMMIFC) was utilized to generate high-resolution FC matrices across 1-20 Hz, which were subsequently processed by a 3D convolutional neural network (3D-CNN) based on a modified VGG architecture. The proposed model achieved classification accuracies of 90.77 % for AD vs HC and 90.38 % for FTD vs HC, with sensitivity and specificity of 88.89 % and 93.10 % for AD vs HC, and 86.96 % and 93.10 % for FTD vs HC, respectively. Beyond classification performance, the analysis identified distinct EEG-based biomarkers within the default mode network. In AD analysis, global efficiency, diffusion efficiency, and clustering coefficient were consistently reduced in the delta and theta bands, reflecting disrupted low-frequency network integration and the theta band presents the most prominent group differences between AD and HC. In FTD analysis, graph theory metrics were also reduced in the delta and theta bands, with the delta band showing the most pronounced group differences compared to HC.