Interpretable Seizure Classification Using Unprocessed EEG With Multi-Channel Attentive Feature Fusion

Identification of seizure type plays a vital role during clinical diagnosis and treatment of epilepsy. However, the clinical evaluation of seizure type is highly dependent on the observed medical symptoms and the experience of the epileptologists who perform the evaluation. A key diagnostic tool is the electroencephalogram (EEG), which captures brain activity and can be used to determine the type of seizure occurring. EEG channels show non-stationary and dynamic behavior following the onset of a seizure event, and each EEG channel can display unique characteristics based on the seizure type and the epileptic foci. This paper proposes a novel deep learning architecture with attention-driven data fusion using raw scalp EEG data from a 10-20 layout, where independent shallow deep networks are trained on each channel. Unlike most state-of-the-art methods that first employ a data engineering step, we directly pass the EEG signal from each channel through a deep convolutional network consisting of SincNet and Conv1D layers, which learn robust features directly from the input signals, increasing model interpretability. However, the importance of each channel and the temporal information varies based on conditions particular to the recording, and this can adversely affect the overall recognition. We propose an approach based on the attentive fusion of channels to ensure only salient features from individual channel encoders are captured, passing the fused information to a Deep Neural Network for classification. Our proposed method has obtained an average F1-score of 0.967 on the Temple University Hospital Seizure Corpus, the largest publicly available seizure dataset.

Automated summary

Accurate identification of seizure types is crucial in the diagnosis and treatment of epilepsy. This paper proposes a novel deep learning architecture with attention-driven data fusion for EEG signal analysis, achieving high performance and interpretability. Our method outperforms state-of-the-art techniques in seizure classification on a large public dataset.