A lightweight convolutional neural network for assessing an EEG risk marker for sudden unexpected death in epilepsy.

BACKGROUND: Convolutional neural network (CNN) has achieved state-of-art performance in many electroencephalogram (EEG) related studies. However, the application of CNN in prediction of risk factors for sudden unexpected death in epilepsy (SUDEP) remains as an underexplored area. It is unclear how the trade-off between computation cost and prediction power varies with changes in the complexity and depth of neural nets.
METHODS: The purpose of this study was to explore the feasibility of using a lightweight CNN to predict SUDEP. A total of 170 patients were included in the analyses. The CNN model was trained using clips with 10-s signals sampled from the original EEG. We implemented Hann function to smooth the raw EEG signal and evaluated its effect by choosing different strength of denoising filter. In addition, we experimented two variations of the proposed model: (1) converting EEG input into an "RGB" format to address EEG channels underlying spatial correlation and (2) incorporating residual network (ResNet) into the bottle neck position of the proposed structure of baseline CNN.
RESULTS: The proposed baseline CNN model with lightweight architecture achieved the best AUC of 0.72. A moderate noise removal step facilitated the training of CNN model by ensuring stability of performance. We did not observe further improvement in model's accuracy by increasing the strength of denoising filter.
CONCLUSION: Post-seizure slow activity in EEG is a potential marker for SUDEP, our proposed lightweight architecture of CNN achieved satisfying trade-off between efficiently identifying such biomarker and computational cost. It also has a flexible interface to be integrated with different variations in structure leaving room for further improvement of the model's performance in automating EEG signal annotation.