Dimensionality reduction for EEG-based sleep stage detection: comparison of autoencoders, principal component...

Area of Science:

• Computational neuroscience
• Biomedical signal processing
• Machine learning for healthcare

Background:

• Automated sleep stage detection relies on predictive models trained with polysomnographic (PSG) recordings.
• High-dimensional PSG data presents computational challenges for model training and deployment.

Purpose of the Study:

• To evaluate the impact of dimensionality reduction techniques on automated sleep stage detection performance.
• To compare principal component analysis (PCA), factor analysis (FA), and autoencoders (AE) in conjunction with various classifiers.

Main Methods:

• Applied PCA, FA, and AE for dimensionality reduction on PSG data.
• Utilized classifiers including random forests (RF), multilayer perceptron (MLP), and long-short term memory (LSTM) networks.
• Tested methods on the MGH Dataset, incorporating electroencephalography (EEG) and other PSG signals (ECG, EMG, respiration).

Main Results:

• Dimensionality reduction techniques generally maintained or improved classification accuracy for sleep stage detection.
• Autoencoders (AE) preserved model performance, while PCA and FA often enhanced accuracy.
• Reduced computational load was observed across tested dimensionality reduction methods.

Conclusions:

• Dimensionality reduction is effective for optimizing automated sleep stage detection models.
• PCA and FA show promise for improving accuracy, while AE offers performance maintenance with reduced complexity.
• These findings support the use of dimensionality reduction to create more efficient and accurate sleep analysis tools.