Comparing Additive and Multiplicative Attention in CNN–LSTM for Ictal Detection

This study comparatively analyzes the performance of six deep learning-based models in epileptic seizure detection. In particular, the effects of additive (Bahdanau) and multiplicative (Luong) attention mechanisms, integrated into the CNN-LSTM backbone, are evaluated in detail. EEG signals were divided into 256 sample-length windows with a stride value of 256, and the training, validation, and test sets were separated by 70, 15, and 15 percent, respectively. Z-score standardization was calculated only with the training statistics and applied to the other subsets. The Bahdanau attention mechanism added to the CNN-LSTM structure achieved the highest performance, with an area under the curve (AUPRC) value of 0.9884, receiver operating characteristic (AUROC) value of 0.9876, and F1 score of 0.9436. The CNN-LSTM model, utilizing the Luong attention mechanism, achieved a similarly high level of success with an AUPRC value of 0.9857 and an AUROC value of 0.9849. The results reveal that attentional mechanisms enhance the model's ability to focus on critical temporal patterns in EEG signals, thereby improving both sensitivity and overall accuracy in seizure detection. These findings suggest that the addition of an attention mechanism to the LSTM and CNN-LSTM architecture offers a practical and reliable improvement in terms of sensitivity-specificity trade-off, training time, and clinical applicability.