Sleep Apnea Detection - Towards Wearables

Sleep Apnea Detection - Towards Wearables

Paper in proceedings (conference paper)

Sleep apnea syndrome (SAS) is a common but serious sleep disorder affecting millions worldwide. It is marked by pauses in breathing over 10 seconds (apnea) or shallow breathing (hypopnea) during sleep, disrupting the sleep cycle and causing fatigue. Untreated SAS increases the risk of stroke, heart disease, high blood pressure, and mental health issues like depression and anxiety. Despite being treatable, many cases remain undiagnosed due to symptoms being misattributed or the inaccessibility of polysomnography (PSG), the diagnostic gold standard. Recently, more accessible methods like electrocardiography or photoplethysmography (PPG) have shown promise, especially with wearable devices. Our work explores using single-channel PPG and deep learning methods (AlexNet, custom CNN, ZF-net) to classify SAS using data from the Multi-Ethnic Study of Atherosclerosis. Using all-night data from 120 patients with the highest apneahypopnea index, we trained on around 9,000 non-apneic and 9,000 SAS-positive intervals, achieving 87.7% accuracy and 85.1% sensitivity, showing the potential for early diagnosis without PSG.

Sleep apnea syndrome (SAS) is a common but serious sleep disorder affecting millions worldwide. It is marked by pauses in breathing over 10 seconds (apnea) or shallow breathing (hypopnea) during sleep, disrupting the sleep cycle and causing fatigue. Untreated SAS increases the risk of stroke, heart disease, high blood pressure, and mental health issues like depression and anxiety. Despite being treatable, many cases remain undiagnosed due to symptoms being misattributed or the inaccessibility of polysomnography (PSG), the diagnostic gold standard. Recently, more accessible methods like electrocardiography or photoplethysmography (PPG) have shown promise, especially with wearable devices. Our work explores using single-channel PPG and deep learning methods (AlexNet, custom CNN, ZF-net) to classify SAS using data from the Multi-Ethnic Study of Atherosclerosis. Using all-night data from 120 patients with the highest apneahypopnea index, we trained on around 9,000 non-apneic and 9,000 SAS-positive intervals, achieving 87.7% accuracy and 85.1% sensitivity, showing the potential for early diagnosis without PSG.

signal processing, deep learning, sleep apnea, wearables,

signal processing, deep learning, sleep apnea, wearables,

KRÁLÍK, M.; NĚMCOVÁ, A.; KOZUMPLÍK, J.; VARGOVÁ, E.

2026

20.12.2024

Computing in Cardiology 2024

Computing in Cardiology

51

United States of America

4