SLEEP-SEE-THROUGH: Explainable Deep Learning for Sleep Event Detection and Quantification From Wearable Somnography

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Science

Background:

• Early diagnosis of sleep disorders is crucial for patient outcomes.
• Multifactorial nocturnal monitoring using wearable devices and AI is a rapidly developing field.
• Current methods for sleep disorder assessment can be invasive or labor-intensive.

Purpose of the Study:

• To develop and evaluate a deep learning model for analyzing sleep patterns from wearable sensor data.
• To assess the model's accuracy in classifying signal quality, breathing patterns, and sleep-related events.
• To enhance model interpretability through saliency maps and confidence indices for clinical translation.

Main Methods:

• Acquisition of optical, air-pressure, and acceleration signals from a chest-worn sensor.
• Elaboration of sensor data into five somnographic-like signals for deep network input.
• Training a deep network on manually labeled data from 20 healthy subjects for three classification tasks: signal quality, breathing patterns, and sleep patterns.

Main Results:

• The deep learning network achieved high accuracy (0.96) in distinguishing normal from corrupted signals.
• Breathing patterns were predicted with high accuracy (0.93), with apnea detection at 0.97.
• Sleep pattern classification showed lower accuracy (0.76), with challenges distinguishing snoring (0.73) from noise (0.61).

Conclusions:

• Deep learning models can effectively analyze multifactorial nocturnal monitoring data for sleep disorder assessment.
• Explainability features like confidence indices and saliency maps improve prediction interpretation.
• This AI-driven approach represents a significant step towards the clinical translation of sleep disorder detection tools.