基于心电图的人工智能用于阻塞性睡眠呼吸暂停检测：一项机制导向的诊断性荟萃分析

文献信息

中文摘要

目的： 基于人工智能（AI）的方法在利用基于心电图（ECG）的生理信号检测阻塞性睡眠呼吸暂停（OSA）方面已展现出良好的性能。然而，现有研究主要关注诊断准确性，而驱动基于AI检测的潜在生理机制在很大程度上仍未得到探索。本研究旨在评估片段级AI模型的诊断性能，并对其生理基础提供机制导向的解读。
方法： 按照PRISMA 2020和PRISMA-DTA指南进行了系统综述和诊断性荟萃分析。系统检索了PubMed、Scopus、Web of Science和IEEE Xplore数据库。纳入使用AI模型基于生理信号进行片段级OSA检测的研究。提取敏感度和特异度值，并使用标准化方法重建2×2列联表，以实现跨研究的可比性。应用双变量随机效应模型，并生成汇总受试者工作特征（SROC）曲线。
结果： 13项研究符合纳入标准。汇总敏感度和特异度分别为0.88和0.89，表明诊断性能较高。SROC曲线显示出极佳的判别能力（AUC > 0.90），研究间异质性为中等。值得注意的是，在不同模型架构中表现一致，提示AI模型依赖于稳定的生理信号模式，而非数据集特异性特征。
结论： 使用生理信号的AI模型在片段级OSA检测中实现了高诊断准确性。重要的是，这些发现提示AI系统可能并非直接检测气道阻塞，而是可能识别下游生理反应，特别是反映在心血管信号中的自主神经系统改变。

英文摘要

OBJECTIVE: Artificial intelligence (AI)-based approaches have shown promising performance in the detection of obstructive sleep apnea (OSA) using electrocardiography (ECG)-based physiological signals. However, existing studies primarily focus on diagnostic accuracy, while the underlying physiological mechanisms driving AI-based detection remain largely unexplored. This study aimed to evaluate the diagnostic performance of segment-level AI models and to provide a mechanism-oriented interpretation of their physiological basis.
METHODS: A systematic review and diagnostic meta-analysis were conducted in accordance with PRISMA 2020 and PRISMA-DTA guidelines. PubMed, Scopus, Web of Science, and IEEE Xplore databases were systematically searched. Studies utilizing AI models for segment-level OSA detection based on physiological signals were included. Sensitivity and specificity values were extracted, and 2 × 2 contingency tables were reconstructed using a standardized approach to enable cross-study comparability. A bivariate random-effects model was applied, and summary receiver operating characteristic (SROC) curves were generated.
RESULTS: Thirteen studies met the inclusion criteria. The pooled sensitivity and specificity were 0.88 and 0.89, respectively, indicating high diagnostic performance. The SROC curve demonstrated excellent discriminative ability (AUC > 0.90), with moderate between-study heterogeneity. Notably, consistent performance across diverse model architectures suggests that AI models rely on stable physiological signal patterns rather than dataset-specific features.
CONCLUSIONS: AI-based models using physiological signals achieve high diagnostic accuracy for segment-level OSA detection. Importantly, these findings suggest that AI systems may not directly detect airway obstruction but instead may identify downstream physiological responses, particularly autonomic nervous system alterations reflected in cardiovascular signals.