Snoring and sleep apnea–a condition where aeroelastic flutter obstructs the airway and stops breathing during sleep–often go hand-in-hand. But diagnosing sleep apnea involves an expensive and time-consuming screening in which the patient has to sleep while monitored by various sensors. To make the process easier, researchers are developing a screening method based only on audio recording.
They started with a pre-trained audio model designed for speech recognition and stripped back computationally-expensive layers that weren’t relevant to snoring. Then they trained the new model using labeled audio data taken from standard clinical testing for sleep apnea. That means the model was told which audio recordings corresponded to “normal” snoring and which showed signs of sleep apnea. From there, the model was able to correctly identify apnea-related audio from fresh recordings just under 74% of the time. While that accuracy isn’t high enough to use the tool for diagnosis, it could help patients pre-screen for sleep apnea at home to decide whether the more invasive testing is warranted. (Image credit: L. Cline; research credit: H. Li et al.; via Physics World)
Leave a Reply