Deep Model Families for EEG-Based Multi-Class Dementia Classification

Deep learning (DL) has shown considerable promise for EEG-based dementia assessment; however, rigorous cross-family comparisons under leakage-free and clinically meaningful evaluation protocols remain limited. To address this gap, we benchmarked twelve DL architectures from four major families - recurrent neural networks (RNNs), temporal convolutional networks (TCNs), Transformers, and self-supervised learning (SSL) - for the multi-class differentiation of Alzheimer's disease (AD), frontotemporal dementia (FTD), and healthy controls (HC). The analysis was conducted on a clinically acquired EEG dataset comprising 36 AD patients, 23 FTD patients, and 29[Formula: see text]HC participants. To prevent subject-level data leakage, we employed strict subject-independent five-fold cross-validation, and to improve clinical relevance, final predictions were aggregated at the subject level using majority voting across EEG segments. Model performance was evaluated using classification metrics, learning stability, statistical ranking, computational efficiency, and interpretability analysis. Among all benchmarked architectures, the self-supervised SimCLR-Time framework achieved the strongest subject-level performance, reaching a mean accuracy of [Formula: see text] and a weighted F1-score of [Formula: see text]. The time-series transformer (TST) was the best-performing fully supervised model, with an accuracy of [Formula: see text], followed closely by the hybrid LSTM-FCN model at [Formula: see text]. In contrast, the EEG-Seq2Seq model showed the weakest performance ([Formula: see text]). Computational profiling further showed that TCN-Attention offered a favorable trade-off between accuracy and edge-oriented efficiency, while SimCLR-Time combined strong diagnostic performance with the fastest inference latency. In addition, Integrated Gradients-based saliency analysis suggested that the top-performing model relied on structured spatiotemporal EEG patterns rather than arbitrary temporal artifacts. Overall, these findings provide a controlled within-dataset benchmark for multi-class dementia classification from raw EEG and suggest that self-supervised and attention-based architectures are promising directions for clinically oriented EEG decision-support systems. However, external validation on larger and more heterogeneous cohorts remains necessary before broader generalization can be claimed.