Microsleep Raw-EEG Classification

Abstract

The ability of univariate convolutional neural networks (1D-CNNs) to classify raw EEG has been demonstrated for various biomedical applications. Using an extensive data set from five driving simulation studies, we investigate whether 1D-CNN is also effective for short-term EEG recorded during microsleep episodes. A standard machine learning solution based on Support Vector Machines (SVM) was developed as a reference methodology. Utilising repeated random crossvalidation, steady convergence with minor fluctuations was observed. It was found that reducing the number of input variables of 1D-CNN using local averaging improved accuracy. Results show that average validation accuracies of 96.3% with SVM and 95.2% with 1D-CNN can be achieved. Sensitivity and specificity revealed insignificant differences.