Key result
Deep learning models on ECG detect obstructive sleep apnea with ~0.80 patient-level AUC.
Why the study?
Most adults with OSA worldwide remain undiagnosed, and although deep learning can automatically extract the autonomic ECG signature of apneic events, a thorough synthesis of these approaches focusing on network architectures, observation duration, and explainability was needed.
Does deep learning applied to electrocardiogram (ECG) signals accurately detect obstructive sleep apnea and assess its severity compared to polysomnography in adults and children suspected of or diagnosed with sleep apnea?
Systematic Review
Does deep learning applied to electrocardiogram (ECG) signals accurately detect obstructive sleep apnea and assess its severity compared to polysomnography in adults and children suspected of or diagnosed with sleep apnea?
Deep learning models applied to ECG signals demonstrate high benchmark accuracy for obstructive sleep apnea detection and AHI estimation, though clinical translation requires broader external validation, patient-level data partitioning, and standardized reporting.
Answers are grounded in this paper and cite their sources.
Requires rigorous external validation before clinical use; extends single-dataset benchmarks while leaving open generalizability and explainability gaps.
Background: It is estimated that there are 936 million adults with OSA worldwide, but most of these individuals are not diagnosed. The electrocardiogram (ECG) has a well-known autonomic signature of apneic events that can be automatically extracted by deep learning. But a thorough synthesis of these approaches is required. Goal: To review peer-reviewed studies for deep learning-based ECG-based OSA detection and severity assessment, focusing on the families of deep learning networks, observation duration, and explainability. Methods: Records were identified by searching the PubMed/MEDLINE, IEEE Xplore, Scopus and Web of Science databases, initially on 15 August 2026 and again on 31 October 2026. Studies utilizing deep learning for apnea/OSA detection, screening, grading or estimation of apnea–hypopnea index (AHI) using ECG were eligible. The assessment of risk of bias was done with a checklist adapted from QUADAS-2 and the synthesis of the evidence was done narratively according to PRISMA 2020. Results: 50 studies (2018-2026) were found. The most common methods were the convolutional neural networks (13/50), CNN–RNN hybrids (9/50), other deep models (9/50), multi-model benchmarks (6/50), and transformers and multimodal fusion (4/50 each). The 60-second epoch was the de facto observation window (23/50). Twelve studies (24%) addressed severity beyond binary classification (8 categorical, 4 continuous AHI); explainability was reported in 10/50 studies, and independent external validation was sparse (7/50). Segment-level accuracy exceeded 98% in several single-dataset evaluations (up to 99.93%) but ranged from 84.7% to 95.8% under stricter designs. At the patient level, the largest clinical-cohort evaluation (11,299 patients; 1,151 in the hold-out test set) reported an AUC of 0.80 (accuracy 73.7%). Continuous AHI estimates correlated with PSG-derived AHI at r = 0.87 and r = 0.90. Conclusions: While deep learning has demonstrated excellent benchmark results for OSA detection based on ECG, the field is limited by reliance on the Apnea-ECG dataset, inconsistent reporting of observation windows, limited external validation, and limited use of explainability. To facilitate clinical translation, standardized reporting of window length, patient-level partitioning, external validation and interpretable behavior should be prioritized.
No takes yet. Share an insight, caveat, or question.
Rejjaoui et al. (2026) conducted a systematic review in Obstructive sleep apnea. Deep learning models for ECG-based obstructive sleep apnea detection and severity assessment vs. Overnight or partial-night polysomnography (PSG) reference standard was evaluated on Obstructive sleep apnea detection and severity assessment performance (segment-level accuracy, patient-level AUC, and AHI correlation). Deep learning models on ECG achieved 84.7%-99.93% segment-level accuracy for OSA detection and an AUC of 0.80 in the largest patient-level clinical cohort.
Synapse has enriched 10 related papers. Consider them for comparative context: