Key Project Outcomes
Many diagnostic sleep tests are carried out at home. However, while these tests have acceptable diagnostic accuracy for obstructive sleep apnea diagnosis in patients with moderate to severe sleep apnea, they tend to underestimate the severity of sleep apnea and can lead to misdiagnosis, in part due to the inability to detect hypopneas terminated by arousals. The goal of our project was to develop a deep learning model for detecting sleep-disordered breathing events (apneas, hypopneas) and arousals from home sleep tests. As part of this goal, we first investigated the relationship between hypopneas and arousals. The American Academy of Sleep Medicine recommends scoring hypopneas when there is a ≥3% oxygen desaturation and/or when the event is associated with an arousal, but there is no rule regarding the duration of the interval between hypopneas and arousals. We found that >90% of hypopnea-associated arousals start within a 20-second range from the end of the hypopnea event (Zitting et al. submitted). This finding helped our model development and could also inform the arousal definition for hypopnea scoring, leading to increased scoring and diagnostic accuracy. We also developed a convolutional Unet model to detect arousals from heart rate using data (Zitting et al. in prep) and are working to extend this model to detect apneas and hypopneas. Our model, trained on >12,000 sleep recordings, performs well against manual arousal scoring (95% accuracy and 0.067 AUCPR score on validation data) and can help improve sleep apnea diagnosis in home sleep tests by detecting hypopneas terminated by arousals, resulting in more patients being identified and treated. We thank the AASM Foundation for their generous support of our project.
A convolutional neural networks model for the detection of cortical arousals from heart rate