DARF-P1DNet: A Dual-View Attention-Guided Residual Fusion Pyramidal 1D Network for EEG-Based Epileptic Seizure Detection

This is the authors' abstract. We don't add key points for this paper.

Seizures cause noticeable abnormal changes in EEG signals. Relying on manual reading of long-term EEG recordings not only costs plenty of time, but it is also easily affected by the doctor's experience and fatigue. To improve the automation and stability of EEG seizure detection, this paper proposes a Dual-view Attention Residual Fusion Pyramidal 1D Network (DARF-P1DNet). First, this method normalizes the single-channel signals from the Bonn EEG dataset. It then constructs the original view and an augmented view for the training phase. Next, two P-1D-CNN encoders are used. They have the same structure, but their parameters are independent. These encoders extract complementary features from the two views. Then, Feature-wise Attention and Residual Fusion are used to filter and keep important information from the dual views. Finally, the model uses a lightweight classification head to perform binary classification between Non-seizure and Seizure. It is trained using weighted cross-entropy loss. In the experiments, we evaluate the model using common classification metrics, ablation studies, and comparison experiments. The overall method has a clear structure. It can fully use the complementary information from the original and augmented signals while keeping the model relatively lightweight. This provides a workable solution for seizure detection in EEG data with small sample sizes.