基于确定学习及心电动力学图的心肌缺血早期检测研究

作者：Sun Qing-Hua; Wang Lei; Wang Cong^*; Wang Qian; Wu Wei-Ming; Zhao Yuan-Yuan; Wang Xi-Ping; Dong Xiao-Nan; Zhou Bin; Tang Min^*

来源：Acta Automatica Sinica, 2020, 46(9): 1908-1926.

DOI：10.16383/j.aas.c190899

• Cardiodynamicsgram (CDG)
• Deterministic learning
• Electrocardiogram (ECG) dataset
• Myocardial ischemia

Early detection of myocardial ischemia is a crucial and challenging problem in cardiovascular disease. In this paper, early detection of myocardial ischemia with normal or nearly normal electrocardiogram (ECG) is investigated by using cardiodynamicsgram (CDG). Firstly, by analyzing the advantages and disadvantages of existing ECG-based machine learning methods for myocardial ischemia detection, a relatively large-scale myocardial ischemia dataset is constructed, which contains ischemic, nearly normal and normal ECGs, as well as coronary stenosis and non-coronary stenosis detected by coronary angiography (CAG). Secondly, for 393 patients in the above dataset with normal or nearly normal ECGs, the deterministic learning algorithm is employed to generate CDGs, and the dynamical features underlying ECGs are extracted. The ischemia of patients is defined by coronary stenosis ≥50%, and the detection model of myocardial ischemia is established using machine learning algorithms. It is shown that ischemia detection model can distinguish myocardial ischemia from non-ischemia effectively. Thirdly, by analyzing the false-positive cases in the above trial, in which each CDG is generated with clear physical meanings, coronary slow flow phenomenons (i.e., non-obstructive coronary lesions) are found in many of the false-positive cases. These cases are re-labeled as ischemia, and a more accurate model for ischemia detection is constructed, with the sensitivity of 90.1%, specificity of 85.2%, accuracy of 89.0% and AUC (area under curve) of 0.93, respectively. As such, in this paper a relatively large-scale myocardial ischemia dataset is constructed which will provide an essential basis for future research on detection algorithms and clinical trials of myocardial ischemia. The established model has the ability to detect ischemia from patients with normal or nearly normal ECGs. Particularly, the CDGs generated by deterministic learning have favorable interpretability, which is helpful for finding the deviations of ischemic data labeling and model errors, and is capable of improving the accuracy of myocardial ischemia detection.

• 出版日期2020
• 单位中国医学科学院北京协和医院; 华南理工大学