Using the electrodermal activity signal and machine learning for diagnosing sleep.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Sleep Medicine

Background:

• Electrodermal activity (EDA) signal use for health diagnostics is growing, driven by machine learning (ML) and wearable devices.
• EDA changes correlate with sleep aspects like sleep stages and sleep-disordered breathing, including obstructive sleep apnoea (OSA).

Purpose of the Study:

• To develop machine learning models for detecting sleep stages and OSA using EDA signals.
• To integrate clinical knowledge of EDA during sleep and OSA with standard statistical features and EDA-specific variables.

Main Methods:

• Supervised machine learning, specifically the extreme gradient boosting (XGBoost) algorithm.
• Utilized EDA signals, incorporating clinical knowledge and EDA-specific variables alongside standard statistical features.

Main Results:

• Achieved an average macro F1-score of 57.5% (five sleep stages) and 66.6% (four sleep stages) for sleep stage detection.
• Obtained 83.7% or 78.4% accuracy for detecting obstructive sleep apnoea (OSA), irrespective of severity, based on the classification measure.

Conclusions:

• Findings support the potential of wearable devices for EDA signal detection in sleep health diagnostics.
• Suggests a future where complete polysomnography (PSG) studies may be supplemented or replaced by wearable-based EDA monitoring for sleep health assessment.