Machine Listening for OSA Diagnosis: A Bayesian Meta-Analysis.

Summary

Across 16 studies (41 models), machine listening using overnight audio achieved pooled sensitivity of 90.3% and specificity of 86.7% for OSA, comparable to common home sleep tests and superior to STOP-Bang screening. Performance was robust across smartphone vs professional microphones; higher sampling frequency and non-contact microphones improved sensitivity.

Key Findings

• Pooled diagnostic accuracy for OSA: sensitivity 90.3%, specificity 86.7%; diagnostic OR 60.8.
• Accuracy comparable between home smartphone recordings and laboratory microphones; deep learning vs traditional ML performed similarly.
• Higher audio sampling frequency and non-contact microphones were associated with increased sensitivity; no evident publication bias.

Clinical Implications

Machine listening can augment or pre-screen OSA referrals, reduce reliance on polysomnography, and expand access via consumer devices; external validation and integration into clinical pathways are needed.

Limitations

• Heterogeneity in datasets, model evaluation (train-test split vs k-fold), and AHI thresholds.
• Potential optimism in train-test split evaluations; limited external validation across diverse populations.

Future Directions

Prospective, device-agnostic external validations; integration trials in primary care to assess referral reduction, cost-effectiveness, and patient-centered outcomes.