JMIR Publications

ABSTRACT

Background:

The traditional method for diagnosing obstructive sleep apnea (OSA) through polysomnography may be expensive and inaccessible. Recent developments in artificial intelligence (AI) proposes the use of photoplethysmography (PPG) to aid OSA diagnosis.

Objective:

This study seeks to evaluate the diagnostic accuracy of an AI-based approach in diagnosing OSA using PPG.

Methods:

PubMed, Embase, Scopus, Web of Science and IEEE Xplore were searched from inception to 20 July 2023. We selected observational studies that evaluated the accuracy of AI-based methods of OSA diagnosis using PPG, compared to conventional sleep testing in adults. Studies focusing solely on the diagnostic accuracy of individual apnoeic events without patient-level classification were excluded. Independent reviewers extracted binary data on diagnostic accuracy, and assessed risk of bias using the Quality Assessment of Diagnostic Accuracy Studies-2 tool. A Bayesian bivariate meta-analysis was used to pool estimates. Further subgroup, sensitivity and meta-regression analyses were conducted.

Results:

From 6931 records, we included 7 studies with 23 AI models, trained on 1272 patients and tested on 757 patients. The overall risk of bias was low. AI trained on PPG achieved a pooled sensitivity of 70.8% (95% Credible Interval (95%CrI): 63.3-77.2%) and specificity of 91.7% (95%CrI: 86.8-95.1%) compared to conventional diagnosis. Bayesian meta-regression identified deep learning (convolutional neural networks) as the AI algorithm with greatest accuracy, with 75.9% sensitivity and 93.5% specificity. The overall specificity improved with increasing categorical AHI severity cut-offs (AHI≥5: 83.8%, AHI≥15: 91.1%, AHI≥30: 92.0%). Additionally, results remained robust to other subgroup meta-regression analyses, including OSA prevalence, country, reference category, feature engineering (domain expert/deep learning), model evaluation (cross-validation/random split), type of sleep study (home sleep apnea testing/polysomnography) and type of device (wearable smart device/sleep study).

Conclusions:

AI trained on PPG have reasonable accuracy and may be considered as a low-cost screening tool. Future work should focus on deep learning to improve the feasibility and accessibility of this approach in primary care. Clinical Trial: PROSPERO (CRD42024534235)