Identificação da condição de sono a partir de características extraídas de sinais de EEG utilizando um módulo Orange Pi Zero

Abstract

This work in progress reports the first results of a project which purpose is the implementation of a microprocessed system for the detection of fatigue, or state of drowsiness, from electroencephalogram (EEG) signals. The use of the Orange Pi Zero platform allows it to have a sleep stage classification system in a compact, mobile and real-time way. So far, the EEG signals used come from a public database, which were acquired from 20 volunteers resulting in 39 signal files with 20 hours of recording. These signal files are divided into epochs (30-second tracks) from which the important frequency bands for the sleep analysis are filtered with the Daubechies’ discrete wavelet transform, and then features are extracted. These features are statistical measures that have as main benefits the reduction of data volume and the highlight of information from the signal, being they the variance, kurtosis and skewness. From these, one can train the random forest classifier to identify whether the patient is awake (W) or sleeping (SLP). The performance of the signal classification is measured by the accuracy calculation, which yields 97% for the test group, corresponding to 40% of the input data (106,285 epochs). Also in this work was made the analysis of the cases where the classifier predicts sleep state with at least one time in advance for a patient and, it was realized that the variance is the most relevant feature for the classifier.