JMIR Publications

ABSTRACT

Background:

Obstructive sleep apnoea (OSA) affects 38% of the population, yet over 90% of cases remain undiagnosed. The current gold standard for OSA diagnosis, polysomnography (PSG), requires specialised equipment, overnight monitoring, and trained personnel, making it inaccessible in primary care and acute settings. With AI advancements, oximetry-based AI models have emerged as a potential alternative for OSA diagnosis.

Objective:

This meta-analysis aims to evaluate the diagnostic accuracy of AI models trained on pulse oximetry readings in diagnosing OSA.

Methods:

A systematic search was conducted across Medline/PubMed, Embase, Scopus, Web of Science, and IEEE Xplore databases. Two blinded independent reviewers screened studies that evaluated the diagnostic accuracy of AI models (including traditional regression and machine learning techniques) trained on SpO₂ recordings, compared to the apnoea-hypopnea index (AHI) as the reference standard. Studies were excluded if they assessed apnoeic events without diagnosing OSA or relied on oxygen desaturation index (ODI) instead of AHI for diagnosis. 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.

Results:

From 6,254 screened articles, 21 studies met the inclusion criteria, encompassing 8,972 participants. No study exhibited a high risk of bias. AI models analysing SpO₂ recordings demonstrated a pooled sensitivity of 91.1% (95% CrI: 89.4–92.7%) and specificity of 86.1% (95% CrI: 82.1–89.4%), with a diagnostic odds ratio (DOR) of 63.3 (95% CrI: 47.3–87.3). The positive likelihood ratio was 6.56 (95% CrI: 5.12–8.51), and the negative likelihood ratio was 0.10 (95% CrI: 0.085–0.122). Among AI models, neural networks achieved the highest sensitivity (93.7%, 95% CrI: 91.2–95.5%) and specificity (89.4%, 95% CrI: 83.3–93.5%), though differences from support vector machines were not statistically significant. Deep learning models showed significantly higher sensitivity than domain expert-based approaches. Sensitivity decreased slightly with higher AHI cut-offs, while specificity increased, though not significantly. There was no substantial publication bias.

Conclusions:

AI models trained on SpO₂ recordings demonstrate high diagnostic accuracy, positioning them as a promising alternative to PSG for OSA diagnosis. This approach offers a more accessible and practical screening or diagnostic method, particularly in primary care and even acute inpatient settings. However, further research is needed to validate its robustness, reliability, and adaptability before widespread clinical implementation.