Awake speech recordings for machine learning diagnosis of obstructive sleep apnea: a Bayesian meta-analysis

Obstructive sleep apnea (OSA) is a prevalent but underdiagnosed condition linked to serious health risks. Due to the limited accessibility of polysomnography (PSG), AI (Artificial Intelligence)-based speech analysis has gained attention as a non-invasive screening tool. This Bayesian meta-analysis evaluates the diagnostic accuracy of AI models trained on awake speech and examines factors affecting performance.

We systematically searched Medline/PubMed, Embase, Scopus, Web of Science, and IEEE Xplore databases. Eligible studies included adults with OSA diagnosis via in-lab polysomnography or home sleep apnea tests and evaluated AI models using speech recordings. Models evaluated using random-split test sets or k-fold cross-validation were included in a Bayesian bivariate meta-analysis and meta-regression. Publication bias was examined using a selection model approach, while risk of bias and evidence quality were assessed with QUADAS-2 and GRADE.

From 6,254 screened articles, 8 studies comprising 24 AI models, trained and tested on 1,060 and 825 participants were included. All studies used professional microphone recordings in the controlled hospital settings. AI models analysing awake speech recordings demonstrated pooled sensitivity and specificity of 82.9% (95% CrI: 80.0-86.4%) and 83.3% (95% CrI: 80.7-86.1%), respectively. The diagnostic odds ratio was 24.3 (95% CrI: 18.2-35.0). Higher mean age improved sensitivity. No significant effects were seen for OSA severity, model type, OSA prevalence, or male percentage. Publication bias was not evident.

AI models trained on awake speech recordings demonstrate good diagnostic accuracy for OSA and hold potential as a practical, scalable screening tool in both clinical and community-based settings.