Graph-informed convolutional autoencoder to classify brain responses during sleep
Sahar Zakeri1, Somayeh Makouei1, Sebelan Danishvar2
1Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran.
Frontiers in Neuroscience
|May 13, 2025
Summary
This study introduces a new machine learning algorithm for classifying sleep states using electroencephalogram (EEG) signals. The robust sleep state (SlS) classifier achieves 99.92% accuracy, improving sleep disorder diagnostics.
Area of Science:
- Biomedical Engineering
- Neuroscience
- Machine Learning
Background:
- Automated machine learning for biomedical signals often struggles with imbalanced datasets.
- Accurate sleep state classification is crucial for diagnosing sleep disorders.
Purpose of the Study:
- To develop a robust sleep state (SlS) classification algorithm using electroencephalogram (EEG) signals.
- To enhance the performance of machine learning models in sleep pattern analysis.
Main Methods:
- Pre-processed EEG recordings from 33 healthy subjects.
- Extracted functional connectivity and recurrence quantification analysis features.
- Developed a novel graph-informed convolutional autoencoder (GICA) with an attention layer.
Main Results:
- Achieved 99.92% accuracy using the SlS-GICA classifier on a significant feature set.
- Identified distinct features differentiating wakefulness, NREM, and REM sleep states (with/without stimuli).
Conclusions:
- The proposed SlS-GICA classifier demonstrates high accuracy and robustness.
- This method holds potential for real-time applications in diagnosing and treating sleep disorders.
