Automatic detection of respiratory events during sleep from Polysomnography data using Layered Hidden Markov Model

Area of Science:

• Medical Technology
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Sleep apnea is a respiratory disorder linked to cardiovascular disease.
• Automatic detection of sleep apnea from polysomnography (PSG) is crucial.
• Detecting subtle events like Respiratory Event Related Arousals (RERAs) remains challenging.

Purpose of the Study:

• To develop an automatic sleep apnea detection system using Hidden Markov Models (HMMs).
• To improve the detection of subtle respiratory events, specifically RERAs.
• To leverage the dynamic modeling capabilities of HMMs for clinical time-series data.

Main Methods:

• Utilized a hierarchical HMM structure, termed Layered HMM, for respiratory event detection.
• Employed features from PSG recordings including airflow, chest/abdomen movements, and blood oxygen saturation (SaO2).
• Analyzed 210 PSG recordings from Massachusetts General Hospital's database.

Main Results:

• Achieved an F1 score between 0.22 ± 0.16 and 0.70 ± 0.08 for event detection across different sleep apnea severities.
• Demonstrated a strong correlation between estimated and PSG-derived Respiratory Disturbance Index (RDI) (R² = 0.91, p < 0.0001).
• Reported a recall of 0.45 ± 0.27 for RERA detection.

Conclusions:

• The layered HMM structure enhances the performance of respiratory event detection during sleep.
• The developed model shows promise for improving the accuracy of sleep apnea diagnosis.
• Further research can refine HMM-based approaches for sleep disorder detection.