Accuracy of deep learning in diagnosis of apnea syndrome: a systematic review and meta-analysis

Alimila Saiyitijiang¹Zhihui Nai²Ping Fan³Ying Gao^3*

• ¹The First Affiliated Hospital of Xinjiang Medical University, Xinjiang, China
• ²Xinjiang Medical University, Urumqi, China
• ³The First Affiliated Hospital of Xinjiang Medical University, Urumqi, China

Objectives This systematic review and meta-analysis was carried out to elucidate the accuracy of image-based deep learning (DL) methods in the real-time detection of obstructive sleep apnea syndrome (OSAS). Methods A systematic search was conducted for studies published since database establishment up to Sep. 25, 2025, across databases including PubMed, Embase, Web of Science, and the Cochrane Library. The included studies were assessed for risk of bias by using the QUADAS-2 tool. During this meta-analysis, a bivariate mixed-effects model was employed and only synthesized the results from the meta-analysis of the validation sets. Meanwhile, subgroup analyses were conducted based on the generation methods of the validation sets. Results A total of 39 original studies were ultimately included, all of which constructed DL images derived from electrocardiogram (ECG) images. Our meta-analysis results suggested that for the comprehensive validation set, the sensitivity, specificity, diagnostic odds ratio (DOR), and the area under summary receiver operating characteristic (SROC) curve were 0.93 (95% CI: 0.90-0.96), 0.95 (95% CI: 0.92-0.96), 252 (95% CI:116-549), and 0.98 (95% CI: 0.42-1.00), respectively. For the independent validation set, the sensitivity, specificity, and SROC curve were 0.93 (95% CI: 0.88-0.96), 0.95 (95% CI: 0.92-0.97), and 0.98 (95% CI: 0.42-1.00), respectively. For the K-fold cross-validation set, the sensitivity, specificity, positive likelihood ratio (LR), and SROC curve were 0.94 (95% CI: 0.88-0.97), 0.94 (95% CI: 0.89-0.96), 15.0 (95% CI: 8.1-27.6) and 0.98 (95% CI: 0.65-1.00), respectively. Conclusions The ECGs-based DL models demonstrate ideal accuracy for the detection of OSAS and appear to be a viable method for real-time detection. During our research process, we found that the modeling was actually based on extracting studies from segments of ECGs, but the extracted segments appeared to vary in duration. Since this aspect was not subjected to subgroup analysis in our study, we plan to conduct further exploration and validation in subsequent research.