Deep neural network detects sleep apnea in sound data

Sleep apnea can be diagnosed with good accuracy by AI analysis of spectrograms—sounds rendered as images—after recording patients’ breathing patterns as they sleep. And the finding may lead to improvements in at-home apnea test devices.

The detection technique was developed and validated by researchers in Japan whose work is running in the American Academy of Sleep Medicine’s Journal of Clinical Sleep Medicine.

In introducing their findings, the authors note that many portable devices for testing sleep apnea at home can’t reliably distinguish between sleep and waking states. Yet tracheal sounds, which can be visualized as a spectrogram, carry information about apnea/hypopnea and sleep/wake status.

To train and test their experimental AI-based detection technique, the researchers used tracheal spectrograms obtained every 60 seconds from more than 1,800 patients as they slept overnight.

The team found a deep neural network with convolutional layers performed well at discriminating breathing status and with reasonable accuracy at discriminating sleep/wake status.

“[W]e demonstrated that tracheal sound spectrograms contain information that can be used for sleep/wake discrimination and apnea/hypopnea detection by deep neural network,” the authors write. “We believe that the analysis of tracheal sound spectrograms using a deep neural network has potential as the basis for an innovative home sleep apnea test device.”

The study is available in full for free.