Downsampling of EEG Signals for Deep Learning-Based Epilepsy Detection

Deep learning-based methods have achieved state-of-the-art accuracy in epileptic seizure detection. However, the high computational demands of deep neural networks pose a significant challenge for implementing epilepsy detection in wearable sensing devices. Existing approaches primarily focus on model lightweighting to reduce the computational burden. This letter, on the other hand, approaches the reduction of inference complexity of deep learning models from a fresh perspective: downsampling of electroencephalogram (EEG) signals. Three types of downsampling methods are presented: direct downsampling, compressed downsampling, and convolutional downsampling. The downsampled EEG signals are directly fed to the deep neural network for seizure detection. Experimental results using the CHB-MIT scalp EEG dataset show that the proposed downsampling methods greatly reduce the computational complexity without sacrificing the detection accuracy. The reduction of computational complexity is nearly proportional to the downsampling factor. In the cases with small to medium downsampling factors, most of the proposed downsampling methods can even improve the seizure detection accuracy.