Yoav N. Nygate, MSc1 • Matt Sprague, MSc1 • Sam Rusk, BSc1 • Chris R. Fernandez, MSc1 • Nathaniel F. Watson, MD, MS2
Introduction
Obstructive sleep apnea (OSA) is an underdiagnosed sleep-related breathing disorder. It is strongly associated with cardiovascular disorders, suggesting a moderate-to-high incidence of co-occurrence among cardiovascular diseases and OSA in patients indicated for multi-night cardiac diagnostic testing. This presents an opportunity to screen for sleep disorders during ambulatory cardiology testing, referring flagged patients for further sleep testing, ultimately increasing the diagnosis and treatment throughput of sleep disorders.
Methods
A Machine Learning (ML) system was developed utilizing over 100,000 diagnostic polysomnography (PSG) studies with concurrently recorded electrocardiogram (ECG) signals. The system leveraged multiple deep neural network models to identify respiratory and sleep-stage-specific ECG patterns, forming an automated tool for ECG-based sleep quality assessment and OSA screening. Clinical validation was performed on a dataset of 185 subjects from a prospective clinical study. PSG results were scored by three RPSGTs, with board-certified sleep physicians providing quality assurance. The ML system’s performance was evaluated against the gold-standard OSA diagnosis using an AHI threshold of 15 events per hour to identify positive OSA cases. Furthermore, to assess ECG-based sleep staging performance, sleep stages were reduced to Wake, Light Sleep (N1 + N2), Deep Sleep (N3), and REM, and agreement was evaluated utilizing an epoch-by-epoch approach.
Results
The ML system achieved a sensitivity and specificity of 90.1% (81.7%, 96.7%) and 84.9% (78.4%, 90.4%) for OSA screening. Furthermore, the ML system produced a sleep staging epoch-by-epoch agreement with a sensitivity and specificity of 91.3% (91.0%, 91.5%) and 95.5% (95.4%, 95.6%) for Wake, 78.7% (78.4%, 78.9%) and 91.2% (91.0%, 91.4%) for Light Sleep, 83.5% (82.7%, 84.3%) and 93.2% (93.1%, 93.3%) for Deep Sleep, and 89.2% (88.7%, 89.6%) and 97.6% (97.5%, 97.6%) for REM.
Conclusion
The ECG-based ML system demonstrated its potential as a scalable solution for automated OSA screening with high sensitivity and specificity in comparison to the gold-standard. The results highlight the ML system’s capability to expand OSA screening, facilitating further diagnosis and treatment of sleep disorders in the broader patient population with cardiovascular disease.
1 EnsoData Research, Ensodata, Madison, WI, USA | 2 Biorhythms Center Integrative Sleep Medicine | 3 Department of Neurology, University of Washington School of Medicine, Seattle, WA, USA
