Multichannel machine learning for polysomnographic diagnosis of obstructive sleep apnea: a Bayesian meta-analysis.

backgroundObstructive sleep apnea (OSA) affects 1 billion people globally, yet over 80% remain undiagnosed. The gold standard for diagnosis, overnight polysomnography, is highly accurate but labor-intensive, leading to delays and increased healthcare burden. Artificial intelligence (AI) models offer a promising alternative. This study pools existing evidence to evaluate AI-based OSA diagnostics.

methodsA systematic search of PubMed, Embase, Scopus, Web of Science, and IEEE Xplore identified studies comparing AI models against the apnea-hypopnea index (AHI) for OSA diagnosis. Studies evaluating models using random-split test sets or k-fold cross-validation were included in a Bayesian bivariate meta-analysis and meta-regression. Risk of bias and evidence quality were assessed using QUADAS-2 and GRADE.

resultsFrom 6,254 records, 7 studies with 19 AI models trained and tested on 7,547 and 7,471 participants were included. No study had a high risk of bias. AI achieved a pooled sensitivity of 89.2% (95% CrI: 81.5-94.3%) and specificity of 87.1% (95% CrI: 82.6-90.8%). Neural networks (NNs) were the best-performing AI model compared to the other subtypes, achieving a sensitivity of 92.8% (95% CrI: 84.8-96.6%) and specificity of 87.8% (95% CrI: 81.1-92.6%). Age and sex had no effect. No publication bias was detected, and the evidence was of high quality.

conclusionNeural Networks AI models trained on polysomnography demonstrated excellent diagnostic accuracy in diagnosing OSA as compared to traditional machine learning. There is a need for further exploration and external validation of AI models to support their integration into routine sleep medicine practice, hence improving access to efficient and accurate OSA diagnosis. PROSPERO registration: CRD42024534235.