SCFNet: An easily transportable epileptiform electrical classification network

There are many obstacles in practical application due to the inconsistent collection standards of EEG signals, especially the models trained on different channel numbers, which are difficult to transfer and use. In order to solve this problem, a neural network architecture with back-end fusion of the single channel feature (SCF) model is proposed. The feature extractor of this model is a CRNN network with single channel input, which is composed of a 1D Inception Layer and ResNet Layer. Since it does not depend on other channel information, it can be more easily migrated to the data of other channel numbers, and the training efficiency is improved due to its parameter reuse. The accuracy of EEG-SCFNet on the IIIC-Seizure dataset was 67.34%. The accuracy rate was 65.88% when the data was migrated from 16 channels to 8 channels, and the training rounds were significantly reduced when the modified network EEG-SCFNet and IIT-SCFNet were used.