Artificial intelligence (AI) approaches applied to physiological signals have shown promise for automated detection of obstructive sleep apnea (OSA). Prior research predominantly reports diagnostic accuracy metrics, but the physiological mechanisms through which AI models discriminate OSA segments remain underexplored. This meta-analysis aimed to quantify diagnostic performance of segment-level AI models using physiological signals—primarily electrocardiography (ECG)-derived measures—and to provide a mechanism-oriented interpretation of what the models are likely detecting.
The study was conducted as a systematic review and diagnostic meta-analysis in accordance with PRISMA 2020 and PRISMA-DTA guidelines. Systematic searches were performed in PubMed, Scopus, Web of Science, and IEEE Xplore to identify studies that applied AI models for segment-level detection of OSA based on physiological signals.
Sensitivity and specificity values were extracted from included reports. To enable cross-study comparability, the investigators reconstructed 2 × 2 contingency tables using a standardized approach. A bivariate random-effects model was applied to pooled data and summary receiver operating characteristic (SROC) curves were generated to assess discriminative ability across studies.
Thirteen studies met the predefined inclusion criteria and were included in the pooled diagnostic analysis. The focus was specifically on segment-level classification tasks using physiological inputs rather than full-night diagnostic labels. Extracted diagnostic measures were harmonized via reconstructed contingency tables, allowing the use of a common statistical framework (bivariate random-effects) for pooled sensitivity and specificity estimation and SROC curve generation.
Across the 13 included studies, pooled sensitivity was 0.88 and pooled specificity was 0.89, indicating high overall diagnostic performance for segment-level detection. The SROC analysis demonstrated excellent discriminative ability, with an area under the curve reported as greater than 0.90.
The authors reported moderate between-study heterogeneity. Despite variability in model architectures and study designs, the pooled performance metrics remained strong. Notably, the analysis found consistent performance across diverse AI architectures, suggesting that the classifiers were leveraging stable physiological patterns rather than idiosyncratic, dataset-specific features.
A central interpretive finding of this meta-analysis is that high-performing ECG-based AI models may not be directly sensing mechanical airway obstruction. Instead, the evidence supports the view that AI systems identify downstream physiological responses to apneic events. Chief among these are changes in autonomic nervous system activity that are reflected in cardiovascular signals such as heart rate and heart rate variability.
The implication is that AI classifiers trained on ECG or related cardiovascular signals are likely detecting autonomic signatures that accompany OSA events—arousal-mediated sympathetic surges, cyclical changes in vagal tone, or other cardiovascular responses—rather than imaging or airflow-derived indices of upper-airway collapse. This mechanism-oriented perspective helps reconcile good diagnostic performance with the indirect nature of the physiological inputs used.
Strengths of the meta-analysis include adherence to PRISMA reporting standards, a standardized reconstruction of diagnostic contingency data to allow pooled analysis, and explicit attention to physiological interpretation rather than solely reporting accuracy metrics.
Limitations reported in the source include moderate heterogeneity across studies and the inherent limitation that ECG-derived or cardiovascular signals capture downstream effects rather than direct evidence of airway obstruction. The analysis also notes that while consistent performance across architectures suggests generalizable signal patterns, the included studies varied in architecture, datasets, and preprocessing steps—factors that can affect external validity.
For clinical translation, the findings support the potential role of ECG-based AI as a screening or triage tool for segment-level detection of OSA-related physiology. However, because the models likely identify autonomic responses rather than airway mechanics, clinicians and developers should be cautious about interpreting positive detections as direct evidence of upper-airway collapse. Further work to link AI-detected physiological patterns with clinical outcomes and polysomnography-confirmed events would strengthen clinical applicability.
AI-based models using physiological signals—particularly ECG and derived heart rate variability measures—achieve high diagnostic accuracy for segment-level detection of obstructive sleep apnea, with pooled sensitivity of 0.88 and specificity of 0.89 and SROC AUC > 0.90. The mechanism-oriented analysis suggests these systems detect downstream autonomic nervous system alterations reflected in cardiovascular signals rather than directly sensing airway obstruction. The study used PRISMA-guided methods and included 13 eligible studies. Between-study heterogeneity was moderate, and authors declare no competing interests. Institutional ethical approval was not required because the analysis used publicly available data.
References and indexing in the original source highlight relevant terms including Artificial intelligence, Electrocardiography, Heart rate variability, Autonomic nervous system, and Sensitivity and Specificity. The DOI and PubMed identifier were reported in the source (PMID 42747635; DOI 10.1007/s11325-026-03798-6).