End-to-End Diffusion Modeling for Clinical EEG Abnormality Detection with Spatial Filtering and Attention

Automated interpretation of clinical electroencephalography (EEG) is challenging due to signal heterogeneity, noise contamination, and inter-subject variability. We propose DiffSA-EEG, a diffusion-based EEG classification framework that integrates learnable spatial filtering, stacked denoising autoencoders (SDA), and convolutional block attention modules (CBAM) within an end-to-end discriminative pipeline. Unlike prior diffusion-based EEG studies that focus on data generation or augmentation, our framework leverages denoising diffusion probabilistic models (DDPMs) directly for discriminative classification as a feature regularizer. We evaluate DiffSA-EEG on two large-scale clinical EEG corpora: the Temple University Hospital Abnormal EEG Corpus (TUAB) and the Temple University Epilepsy Corpus (TUEP). DiffSA-EEG consistently outperforms established baselines—including EEGNet, Deep4Net, ChronoNet, temporal convolutional networks, and the Diff-E backbone—across accuracy, AUC-ROC, and AUC-PR. Ablation analyses reveal that optimal component combinations are dataset-dependent: spatial filtering with SDA is most effective for TUAB, while SDA with CBAM yields superior performance on TUEP. Grad-CAM-based interpretability analysis further shows that the model captures clinically plausible spatial patterns aligned with established neurophysiological biomarkers.