A Temporal-Spectral Based Single-lead Electroencephalogram Feature Fusion Network may Provide Potential Clinical Biomarker for Cardiac Arrest

Cardiac arrest is a fatal condition requiring rapid identification and intervention. Our team “SHE Lab” develops a deep neural network for automated detection from single-lead electroencephalogram (EEG) as part of the 'Predicting Neurological Recovery from Coma After Cardiac Arrest: The George B. Moody PhysioNet Challenge 2023
^,
. Our model comprises complementary time-domain and spectral-domain to extract prognostic biomarkers. The adaptive time-domain convolution block directly analyzes the EEG waveform. The multi-resolution wavelet decomposition block captures discriminative spectral bands. Feature fusion integrates this multi-modal information before final classification. While our team was unable to be scored on the test set, experiments demonstrate good performance with accuracy 78.1%, AUROC 0.914, AUPRC 0.942, F1-score 0.841 on our held-out subset of the training set. Compared to methods based on multi-lead EEG, our automated single-lead interpretation model can achieve accessible and scalable monitoring, providing a powerful and universal method to explore the predictive function of EEG. The proposed biomarkers demonstrate the low-cost, rapid diagnosis, real-time care in clinical practice. Therefore, the biomarkers may provide important value for the prognosis evaluation and timely treatment of patients with cardiac arrest.