SeriesSleepNet: an EEG time series model with partial data augmentation for automatic sleep stage scoring.
Minji Lee1, Heon-Gyu Kwak2, Hyeong-Jin Kim3
1Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, Republic of Korea.
Frontiers in Physiology
|September 13, 2023
Summary
SeriesSleepNet, a novel deep learning model using convolutional neural networks (CNN) and bidirectional long short-term memory (bi-LSTM), accurately scores sleep stages from electroencephalography signals. This advanced model enhances sleep disorder diagnosis and sleep quality assessment.
Area of Science:
- * Neuroscience
- * Biomedical Engineering
- * Machine Learning
Background:
- * Accurate sleep stage scoring is crucial for diagnosing sleep disorders and assessing sleep quality.
- * Traditional methods often rely on manual annotation or simpler algorithms, which can be time-consuming and less precise.
- * Automatic sleep stage scoring using electroencephalography (EEG) signals offers a more efficient and objective approach.
Purpose of the Study:
- * To develop and validate an automatic sleep stage scoring model named SeriesSleepNet.
- * To leverage convolutional neural networks (CNN) and bidirectional long short-term memory (bi-LSTM) networks for enhanced time-series analysis of EEG data.
- * To improve the accuracy and performance of sleep stage classification using partial data augmentation and a novel loss function.
Main Methods:
- * Implemented SeriesSleepNet, a hybrid deep learning architecture combining CNN for intra-epoch feature extraction and bi-LSTM for inter-epoch sequential learning.
- * Utilized single-channel raw electroencephalography (EEG) signals as input.
- * Applied partial data augmentation to capture temporal dependencies within short time series.
- * Employed a custom loss function for model fine-tuning and validated on the Sleep-EDF and SHHS datasets.
Main Results:
- * Achieved high performance in five-class sleep stage classification, with overall accuracies of 0.87 (Sleep-EDF) and 0.84 (SHHS).
- * Demonstrated superior performance compared to baseline models and state-of-the-art methods, evidenced by overall F1-scores of 0.80 and 0.78, and kappa values of 0.81 and 0.78.
- * The CNN and bi-LSTM components, along with partial data augmentation, significantly contributed to the model's effectiveness.
Conclusions:
- * SeriesSleepNet provides a highly accurate and robust framework for automatic sleep stage scoring using single-channel EEG.
- * The model's performance suggests its potential utility in clinical settings for sleep disorder diagnosis and sleep quality assessment.
- * The proposed architecture effectively captures both local and sequential temporal information in EEG signals for improved classification.
