A dual-branch time-frequency fusion network for EEG-based classification of Alzheimer's disease and frontotemporal
Xiaoli Yang1, Xiao Li1, Chenchen Wang1
1School of Medical Technology and Engineering, Henan University of Science and Technology, Luoyang, China.
Abstract:
BackgroundAlzheimer's disease (AD) and frontotemporal dementia (FTD) exhibit substantial overlap in clinical manifestations and patterns of brain functional degeneration, which poses significant challenges for automated classification based on electroencephalography (EEG).ObjectiveThis study aims to develop an EEG-based framework capable of simultaneously capturing temporal dynamics and frequency-related characteristics of EEG signals for discrimination among AD, FTD, and cognitively normal (CN) subjects.MethodsA Dual-Branch Time-Frequency Fusion Network (DBTF-Net) based on routine clinical resting-state EEG recordings acquired under eyes-closed conditions is proposed. The model employs parallel temporal and frequency branches to process raw EEG time-series signals and their corresponding time-frequency representations. A global temporal dependency construction mechanism is introduced in the temporal branch to capture both local temporal patterns and long-range temporal dependencies. Feature-level fusion is then performed across the two branches to achieve a collaborative representation of multidimensional brain functional information. The proposed method was systematically evaluated on one three-class classification task (AD versus FTD versus CN) and multiple binary classification tasks.ResultsExperimental results from five-fold cross-validation at the epoch level show the classification accuracies of DBTF-Net as 86.36%±4.28%, 83.01%±6.15%, 92.13%±10.35%, and 88.74%±7.69%% for AD versus FTD versus CN, AD versus CN, FTD versus CN, and AD versus FTD, respectively.ConclusionsThe proposed DBTF-Net leverages temporal and time-frequency information in EEG signals and provides classification of AD and FTD. Visualization analysis further indicates that the model attends to disease-relevant discriminative patterns in time-frequency representations, enhancing the interpretability of its classification decisions.