Zheng Bo, Liu Yaokun, Zhang Jingyi, Ma Terry T, Zhou Yun, Chen Yongkai, Yang Ying, Ma Wei, Fan Fangfang, Jia Jia, Zhang Yan, Li Jianping, Zhong Wenxuan
Department of Cardiology, Peking University First Hospital, Beijing, China.
Institute of Cardiovascular Disease, Peking University First Hospital, Beijing, China.
Intern Emerg Med. 2025 May 21. doi: 10.1007/s11739-025-03968-6.
Coronary functional assessment plays a critical role in guiding decisions regarding coronary revascularization. Traditional methods for evaluating functional myocardial ischemia, such as invasive procedures or those involving radiation, have their limitations. Echocardiographic myocardial strain has emerged as a non-invasive and convenient indicator. However, the interpretation of strain values can be subject to inter-operator variability. Artificial intelligence (AI) and machine learning techniques may promise to reduce the variability. By training AI algorithms on a diverse range of echocardiographic data, including strain values, and correlating them with ischemia, it may be possible to develop a robust and automated diagnostic tool. This study aims to provide a non-invasive and effective solution for automated myocardial ischemia detection that can be used in clinical practice. To construct the machine learning model, we used an automatic left ventricular endocardium tracing tool to extract myocardial strain data and integrated it with six clinical features. A coronary angiography-derived fractional flow reserve (caFFR) ≤ 0.80 was defined as the indicator of myocardial ischemia. A total of 636 suspected coronary artery disease subjects were enrolled in this pilot study, where 282 cases (44.3%) had myocardial ischemia. These subjects were randomly divided into training (n = 508) and testing (n = 128) sets at a 4:1. Using ensemble-learning algorithms to train and optimize the model, its diagnostic performance versus caFFR was diagnostic accuracy 85.9%, sensitivity 88.9%, specificity 83.1%, positive predictive value 83.6%, negative predictive value 88.5%. The optimized model achieved an area under the receiver operating characteristic curve (AUC) of 0.915 (95% confidence interval [CI] 0.862-0.968). Our machine learning prototype model based on echocardiographic myocardial strain shows promising results in detecting myocardial ischemia. Further studies are needed to validate its robustness and generalizability on larger patient populations.
冠状动脉功能评估在指导冠状动脉血运重建决策中起着关键作用。评估功能性心肌缺血的传统方法，如侵入性检查或涉及辐射的检查，都有其局限性。超声心动图心肌应变已成为一种非侵入性且便捷的指标。然而，应变值的解读可能会受到操作者间差异的影响。人工智能（AI）和机器学习技术有望减少这种差异。通过在包括应变值在内的各种超声心动图数据上训练AI算法，并将其与缺血情况相关联，有可能开发出一种强大的自动化诊断工具。本研究旨在提供一种可用于临床实践的非侵入性且有效的自动化心肌缺血检测解决方案。为构建机器学习模型，我们使用自动左心室心内膜追踪工具提取心肌应变数据，并将其与六个临床特征相结合。将冠状动脉造影衍生的血流储备分数（caFFR）≤0.80定义为心肌缺血的指标。本初步研究共纳入636例疑似冠状动脉疾病患者，其中282例（44.3%）有心肌缺血。这些患者按4:1随机分为训练组（n = 508）和测试组（n = 128）。使用集成学习算法训练和优化模型，其与caFFR相比的诊断性能为诊断准确率85.9%，敏感性88.9%，特异性83.1%，阳性预测值83.6%，阴性预测值88.5%。优化后的模型在受试者工作特征曲线（AUC）下的面积为0.915（95%置信区间[CI] 0.862 - 0.968）。我们基于超声心动图心肌应变的机器学习原型模型在检测心肌缺血方面显示出有前景的结果。需要进一步研究以验证其在更大患者群体中的稳健性和通用性。
