Bridging data scarcity and explainability in EEG based Alzheimer’s prediction using cGAN augmented GATv2-LSTM networks

Abstract Accurate and early diagnosis of Alzheimer’s Disease (AD) is still a key factor in the treatment of the disease and patient health. The proposed model for the current work is an EEG-based multi-classification model which is tested on two separate datasets with different diagnostic classes (Dataset I: Healthy Control (HC), Mild Cognitive Impairment (MCI), Alzheimer Disease (AD); Dataset II: Healthy Control (HC), Fronto-temporal Dementia (FTD), Alzheimer Disease (AD)). A Conditional GAN (cGAN) data augmentation technique was used to tackle the issue of class imbalance, especially MCI, FTD and HC were underrepresented. Analysis of the fidelity of the generated synthetic EEG signals was conducted using the Wasserstein distance and Kernel Density Estimation (KDE). The results reveal that there is a high similarity between real and synthetic signals for Dataset I in frontal channels (Fp1, F3) and with larger channel-specific deviations for the MCI class. Whereas, for Dataset II, both FTD and HC data showed good overall similarity between the real and synthetic EEG signals, with some larger channel-specific deviations observed for the FTD class. The proposed Hybrid GATv2-LSTM model integrates both spatial and temporal information from EEG signals for Alzheimer’s disease classification. Specifically, Graph Attention Network v2 (GATv2) learns the spatial relationships among EEG channels by modeling their connectivity, while Long Short-Term Memory (LSTM) captures the temporal dynamics of brain activity. Furthermore, adaptive attention fusion and residual connections are incorporated to effectively combine spatial and temporal features, enhance feature learning, and improve classification performance. Under a sample-level 80/20 split, it achieved the classification accuracies of 94.60% on Dataset I and 89.19% on Dataset II, better than the model without data-augmentation and standalone model. To improve the interpretation of the model, Integrated Gradients (IG) was employed to evaluate the contribution of individual EEG channels to the classification outcomes. The results demonstrated that the channels over the frontal, parietal, central, and occipital areas were the most relevant for the model, while the channels over other areas had a relatively small impact. This finding is consistent with previous studies on EEG-based Alzheimer’s detection. This proof-of-concept study is a promising exploration of the feasibility of using validated synthetic EEG data and interpretable deep learning models to achieve data efficient EEG-based AD classification.