Disclosures: Cho reports receiving a licensed national patent related to this study. Please see the study for relevant financial information from all other authors.
Sleep breath sounds recorded via a smartphone gave an accurate prediction of obstructive sleep apnea, researchers report in JAMA Otorhinolaryngology – Head & Neck surgery.
“The gold standard diagnostic method for OSA is assisted, all-night, laboratory-based polysomnography, which involves recording numerous physiological signals that are manually scored by sleep technicians or board-certified physicians. Therefore, laboratory polysomnography is expensive and accessibility to a sleep facility is not always easy,” Sung-Woo Cho, MD, from the Department of Otorhinolaryngology – Head and Neck Surgery at Bundang Hospital, Seoul National University, Seoul National University College of Medicine, Seongnam, South Korea, and its colleagues wrote. “Given the high prevalence of OSA, performing polysomnography in the lab overnight may not be practical for all patients.”
The cross-sectional study recruited 423 patients (mean age, 48.1 years; 84.1% male) who visited the Bundang Hospital Sleep Center of Seoul National University for snoring or sleep apnea from August 2015 to August 2019. Patients recorded sleep audio with a smartphone during routine polysomnography in the laboratory overnight. The researchers performed binary classifications for different threshold criteria based on an apnea-hypopnea index threshold of five, 15, or 30 events per hour and created four regression models, including noise reduction without feature selection , noise reduction with feature selection, neither noise reduction nor feature selection, and the feature selection without noise reduction.
The researchers divided the data into training (n=256) and test (n=167) datasets and patients were grouped into normal OSA (n=43), mild OSA (n=80), OSA moderate (n=109) or severe OSA (n=191).
Smartphone-recorded audio yielded 88.2% accuracy for an apnea-hypopnea index threshold of five events per hour, 82.3% for 15 events per hour, and 81.7% for 30 events per hour. hour. The areas under the curve were 0.9, 0.89, and 0.9 for five, 15, and 30 events per hour, respectively.
The four regression models showed similar results with correlation coefficients ranging from 0.77 to 0.78. Smartphone recorded audio that was not denoised and had only selected attributes yielded the highest correlation coefficient in the regression analysis (r = 0.78). The apnea-hypopnea index (beta = 0.33) and sleep efficiency (beta = -0.2) were associated with OSA estimation error.
“These prediction models gave good prediction performance, and we found that noise suppression was not mandatory for good prediction performance. … Future research should be extended to incorporate real smartphone recordings at home using various smartphones,” the researchers wrote.