EnsembleNet: Single-Channel EEG Sleep Stage Classification based on Ensemble Architecture of Deep Convolutional Neural Networks

Abstract

Sleep disorders can significantly affect health, safety, and quality of life, making sleep stage classification essential for diagnosis and treatment. While deep learning-based automated classification has shown promise, multi-class classification with a single convolutional neural network (CNN) can be overly complex and prone to overfitting. Therefore, we propose a novel architecture, EnsembleNet, using a single electroencephalogram (EEG) channel. Our approach consists of an ensemble of per-epoch CNN-based binary classifiers for feature extraction, followed by an ensemble of per-sequence bidirectional long short-term memory (BiLSTM) models to capture temporal dependencies. A shared, relatively simple architecture for binary classifiers helps mitigate overfitting, while a two-step training approach was chosen to reduce computational workload by separately training CNN-based binary classifiers and BiLSTM models. Model performances were evaluated using k-fold cross validation on Sleep-EDF-20 (k = 20) and Sleep-EDF-78 (k = 10) datasets where an EEG signal is either Fpz-Cz or Pz-Oz channel. On Sleep-EDF-20, the model achieved F1-scores of 86.94% − 90.16% and 83.16% − 89.50% (non-N1), and 55.32% and 46.14% (N1) with overall accuracies of 87.07% and 85.18%, for the Fpz-Cz and Pz-Oz channels, respectively. Similarly, on Sleep-EDF-78, F1-scores reached 83.35% − 94.06% and 71.94% − 92.15% (non-N1), and 53.17% and 48.51% (N1) with overall accuracies of 85.95% and 82.86%, for the Fpz-Cz and Pz-Oz channels, respectively. Our model outperforms selected state-of-the-art methods, offering an efficient alternative to complex multi-class CNN architectures. Additionally, the ensemble of binary classifiers enhances explainability and facilitates performance optimization through systematic assessment.