Moderate evidence
More
Bedroom audio alone tracked nighttime oxygen drops in a study of 802 test nights
Study title: Contactless Oxygen Desaturation Prediction and Oxygen Desaturation Index (ODI) Estimation from Nocturnal Breathing Sounds via Deep Learning
- Evidence type
- Controlled or validated comparison
- Sample
- n = 802
- Sources
- 1 sourcePeer-reviewed
- Published
- 1 min read
- Truvace record
- TRV-2026-2959cite or check it
Study title: Contactless Oxygen Desaturation Prediction and Oxygen Desaturation Index (ODI) Estimation from Nocturnal Breathing Sounds via Deep Learning
Researchers built a deep learning model that listens to nocturnal breathing sounds to detect oxygen desaturation events and estimate the oxygen desaturation index (ODI), without any sensor touching the sleeper. On the independent test set it reached 0.79 sensitivity, 0.95 specificity and 0.92 accuracy at the 30-second epoch level, and its ODI estimates tracked polysomnography values at r = 0.93.
Read the study abstract
Study objectives Intermittent hypoxemia is increasingly recognized as a stronger predictor of cardiovascular morbidity and mortality in obstructive sleep apnea (OSA) than the apnea-hypopnea index (AHI). We developed a sound-based AI model to detect oxygen desaturation events and estimate the oxygen desaturation index (ODI) from nocturnal breathing sounds in a fully contactless manner. Methods The model was trained using 2,973 nights of synchronized polysomnography (PSG) and audio recordings obtained from hospita…

The gain
In a study testing a breathing-sound AI model on 802 independent nights against polysomnography, it detected oxygen desaturation epochs with 0.79 sensitivity and 0.95 specificity, and its estimated ODI correlated with the polysomnography value at r = 0.93.
In brief
The work matters because intermittent hypoxemia is increasingly seen as a stronger predictor of cardiovascular illness and death in obstructive sleep apnea than the apnea-hypopnea index. A contactless method could make oxygen-related assessment easier outside sleep labs, and the authors describe it as potentially scalable for real-world settings.
Main points
- The model was trained on 2,973 nights of synchronized polysomnography and audio recordings from hospital and home settings.
- Validation used 664 nights and independent testing used 802 nights.
- Audio was converted into Mel spectrograms and processed by a Vision Transformer-based architecture.
- Epoch-level results: sensitivity 0.79, specificity 0.95, precision 0.78, F1 score 0.79, accuracy 0.92 and Cohen's κ 0.74.
- Estimated ODI matched polysomnography-derived ODI with r = 0.93 and R² = 0.87.
The full story
The team used nights recorded at the same time by polysomnography and audio, in both hospitals and homes, so the model learned from different recording environments. Polysomnography served as the reference for when oxygen desaturation occurred.
Performance was judged two ways: how well it flagged desaturation in each 30-second epoch, and how closely its overall ODI matched the reference, using Pearson correlation, Bland-Altman analysis, mean absolute error and root mean square error.
The authors conclude that the approach enables contactless detection of desaturation events and reliable ODI estimation, and may offer a scalable non-contact way to assess sleep-disordered breathing.
Sources
1 on file- SLEEPdoi:10.1093/sleep/zsag295
Talk it over
The model reached 0.79 sensitivity and 0.95 specificity across 802 test nights. How would you weigh the nights where it missed desaturation epochs against the convenience of a method with no sensor on the sleeper?
Bedroom audio alone tracked nighttime oxygen drops in a study of 802 test nightshttps://truvace.com/item/contactless-oxygen-desaturation-prediction-and-oxygen-desaturation-index-odi-est
The model reached 0.79 sensitivity and 0.95 specificity across 802 test nights. How would you weigh the nights where it missed desaturation epochs against the convenience of a method with no sensor on the sleeper?
Answer in the threadThe model was trained on 2,973 nights from hospitals and homes, and tested on 802 independent nights against polysomnography. What does that design tell us about how it would do for people sleeping in other bedrooms with other noise?
Answer in the threadIf you were someone with suspected sleep apnea who could not get to a sleep lab, what would you want to know about this audio method before relying on its ODI estimate?
Answer in the thread
If you have more time
Look up how oxygen desaturation is measured outside sleep labs today. What would the audio approach need to match to replace it?
Answer in the threadThe estimated ODI correlated with polysomnography at r = 0.93. What kind of follow-up study would show whether it changes treatment decisions for patients?
Answer in the thread
ace
What readers say