Temporal Feature Extraction and Machine Learning for Classification of Sleep Stages Using Telemetry Polysomnography.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Accurate sleep stage detection is vital for diagnosing sleep disorders.
• Polysomnography (PSG) is the gold standard, capturing multiple physiological signals.
• Complex neural networks are often used, but simpler methods are sought.

Purpose of the Study:

• To evaluate simpler machine learning models for sleep staging using PSG data.
• To compare various feature extraction techniques in combination with machine learning classifiers.
• To investigate the efficacy of data augmentation for imbalanced sleep data.

Main Methods:

• Extracted features using power spectral density, Higuchi fractal dimension, SVD entropy, permutation entropy, and DFA.
• Employed machine learning models: XGBoost, Extra Trees, Random Forest, and LightGBM.
• Utilized Synthetic Minority Oversampling Technique (SMOTE) for data augmentation.

Main Results:

• The XGBoost classifier with an ensemble of all features achieved the highest performance.
• Accuracies ranged from 87% (5-stage) to 97% (2-stage).
• Average F1-scores ranged from 84.6% to 93.5% across different sleep stage configurations.

Conclusions:

• A combination of feature extraction measures with XGBoost offers an efficient and high-performing alternative for sleep staging.
• This approach surpasses the performance of many recent deep neural networks.
• The findings suggest a more accessible and efficient avenue for sleep disorder diagnosis and treatment.