ROM-based inference method built on deep learning for sleep stage classification

Abstract

We used a classical deep feedforward neural network (DFFNN) for an automatic sleep stage scoring based on a single-channel EEG signal. We used an open-available dataset, randomly selecting one healthy young adult for both training (≈5%) and evaluation (≈95%). We also augmented the validation by using 5-fold cross validations for the result comparisons. We introduced a new method for inferring the trained network based on a ROM module (memory concept), so it would be faster than directly inferring the trained Deep Neural Network (DNN). The ROM content is filled after the DNN network is trained by the training set and inferred using the testing set. An accuracy of 97% was achieved in inferring the test datasets using ROM when compared to the classic trained DNN inference process.