Sleep Stages Classification Using Music Made from EEG By LSTM Networks

Hamidreza Jalali; Majid  Pouladian; Ali Motie  Nasrabadi; Azin  Movahed

doi:10.18502/fbt.v13i2.21933

Hamidreza Jalali Department of Biomedical Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran
Majid Pouladian Department of Biomedical Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran
Ali Motie Nasrabadi Biomedical Engineering Department, Faculty of Engineering, Shahed University, Tehran, Iran
Azin Movahed Music Department, School of Performing Arts and Music, College of Fine Arts, University of Tehran, Tehran, Iran

DOI: https://doi.org/10.18502/fbt.v13i2.21933

Keywords: Single-Channel Electroencephalogram; Sleep Staging; Electroencephalogram Sonification; Long Short-Term Memory.

Abstract

Purpose: Automatic classification of sleep stages is one of the fundamental factors in diagnosing sleep disorders to prevent and treat various diseases, and it can significantly aid in saving specialists' time and energy. In this study, a new method for converting Electroencephalogram (EEG) signals to music for sleep stages classification is proposed.

Materials and Methods: A total of 15.233, 30-second data segments from the Sleep-EDF database were used as the statistical population for this evaluation. Initially, the performance of Long Short-Term Memory (LSTM) networks for music sequence generation is evaluated with the music database and the best structure is selected. Subsequently, single-channel EEG data are mapped to music pieces using the selected network. Seven features are extracted from the generated music sequences and applied to classification structures.

Results: The selected LSTM structure was able to identify musical sequences with an accuracy of 93.3% of the musical pieces. The overall classification accuracy for the five sleep stages according to the AASM standard is 85.3% for the Sleep-EDF database. Accuracy of classifying W, N1, N2, N3, and REM stages are 86.1%, 77.3%, 95.4%, 96.3%, and 71.4%, respectively. Another objective of this study is to present a novel single-channel EEG sonification method, achieving classification accuracy that is either higher than or comparable to contemporary methods.

Conclusion: The results of this study show that audio signal mapping with LSTM networks contains effective information for sleep stage classification, and the classification accuracy increased by 1% compared to the method of a similar study and by 3% compared to most studies.