Enhancing sleep stage classification with 2-class stratification and permutation-based channel selection

We present a method that uses a convolutional neural network (CNN) called EEGNeX to extract and classify the characteristics of sleep-related waveforms from electroencephalographic (EEG) signals in different stages of sleep. Our results showed that the CNN model with 128 channels achieved high performance, distributing the sleep stages into 2-class models. We used a permutation-based channel selection process and using the top 3 channels, we achieved a performance greater than 80% in accuracy, Fscore, precision, recall, area under the receiver operating characteristic (AUROC) and kappa value, except when classifying N1 versus N2, where the average kappa value was 0.52. Performance is shown to decrease when using the 3 channels recommended by the American Academy of Sleep Medicine (AASM) or 3 random channels. Overall, the results showed that 2-class CNN models with 3 channels selected with a permutation-based approach achieve good performance in the classification of sleep stages from EEG signals, with a computational cost much lower than using 128 EEG channels.