Deep Neural Network method for Sleep Stages Classification using spectrogram of signal based on transfer learning with different domain data

Abstract

Sleep stages Classification is a useful way to diagnose sleep problems. This is based on the processing of bio-signals (ECG, EEG, EOG, PPG). The less complex this signal is, the better the detection and processing. Feature extraction methods using hand are tedious and long lasting. Extraction of features without hand intervention are deep features, which are usually extracted from images. Analysis of time-frequency characteristics of non-static bio-signals is very important and has useful information. In this study, time-frequency image was extracted using ECG signal spectrogram and deep features were extracted using convolutional neural network. After extracting deep features, sleep stages were classified using deep transfer learning method. Network training was performed using one of the ECG signal and testing was performed with the other ECG signal channel.The results show that it is possible to detect sleep stages with acceptable accuracy with different amplitudes of signals. Sleep stages were detected with 98.92% accuracy and 96.52% sensitivity.