ECG‐based cardiodynamicsgram can reflect anomalous functional information in coronary artery disease

Abstract Background The cardiodynamicsgram (CDG), a novel noninvasive method, extracts dynamic ST‐T segment information from an electrocardiogram (ECG) through deterministic learning. Hypothesis The CDG can reflect anomalous functional information in coronary artery disease (CAD). Methods We retrospectively enrolled 456 patients with suspected CAD who underwent coronary computed tomography angiography (CCTA) from January 2020 to 2022, followed immediately by standard 12‐lead ECG acquisition. Positivity for CAD were defined as CCTA ≥ 50% or CT‐derived fractional flow reserve (CT‐FFR) ≤ 0.8. A CDG value <0 was considered negative; otherwise, it was considered positive. We also evaluated the diagnostic performance of the CDG in the ECG‐diagnosis‐negative subgroup and in patients who had undergone invasive coronary angiography (ICA) after CCTA. Results Of 362 patients, 168 (46.41%) were positive for CAD, and 178 (49.17%) were men. The median age was 59 (52−66) years. The accuracy of the CDG in the diagnosis of CAD was 79.56%, with a sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) of 75.60%, 82.99%, and 0.836 (95% CI: 0.794−0.878), respectively. Similarly, in the ECG‐diagnosis–negative subgroup (n = 223), the accuracy of the CDG was 80.27%, with an AUC of 0.842 (95% CI: 0.790−0.895). Among the 11 patients with CAD confirmed by ICA, 10 were diagnosed positive by the CDG. Furthermore, the CDG values and CT‐FFR were correlated (r = −.395; p < .001). Conclusions The ECG‐based CDG has relatively high specificity and accuracy for the diagnosis of CAD and reflects functional cardiac information to some extent. It has the potential to be used as a screening tool for suspected CAD patients before CCTA.

the potential to be used as a screening tool for suspected CAD patients before CCTA.

K E Y W O R D S
cardiodynamicsgram, computed tomography-derived fractional flow reserve, coronary artery disease, coronary computed tomography angiography, electrocardiogram

| INTRODUCTION
Coronary artery disease (CAD) is one of the most common cardiovascular diseases leading to death in the global population. 1 Coronary atherosclerosis is the primary reason for CAD, which manifests as progressive occlusion of the coronary arteries with consistently lower oxygen and nutrient perfusion. 2 Asymptomatic, stable plaques can also become unstable at any time due to plaque rupture or erosion and even directly lead to fatal events. 3 Coronary computed tomography angiography (CCTA) is the preferred diagnostic test for CAD, 4 used to assesses the position and extent of atherosclerotic plaque and the degree of coronary artery stenosis. CCTA has the imaging advantages of more availability and less invasiveness. However, it is radioactive, and the severe allergic reactions and kidney damage caused by contrast injection cannot be ignored. 3,5 Moreover, CCTA provides only anatomical information on the coronary arteries, and coronary stenosis are often overestimated, only a minority of severe stenoses diagnosed by CCTA are finally recognized as myocardial ischemia. 6 To obtain functional information requires complex postprocessing operations on CCTA images, such as CT-derived fractional flow reserve (CT-FFR). The CT-FFR is a noninvasively obtained parameter used to assess hemodynamically significant stenoses without the need for additional testing and radiation exposure, and CT-FFR accuracy of 90% compared to invasive FFR. [7][8][9] Clinically, the electrocardiogram (ECG) is the most common test for patients with suspected CAD. 10 However, interpretation of the ECG signal is highly influenced by the observers, and fluctuations in the ECG signals generated by early CAD are too subtle to detect. 11 Deterministic learning has been proposed for the accurate modeling and rapid recognition of temporal or dynamical patterns. [12][13][14] The cardiodynamicsgram (CDG) is a new noninvasive tool for detecting subtle changes in the ST-T segment of the ECG. The CDG can extract quantitative dynamic information on its spatial and temporal dispersion by deterministic learning and be visualized with a three-dimensional (3D) image to illustrate the abnormalities of the heart. Recently, some studies have confirmed that the CDG can predict ≥50% of coronary stenosis and has exceptional diagnostic performance for myocardial ischemia. 15 However, no study has evaluated the ability of CDG to reflect abnormal cardiac function and thus aid in the screening of CAD.
Precise, easy, rapid, and noninvasive tools for screening CAD in the clinic are urgently needed. Hence, this study attempted to evaluate the diagnostic performance of the CDG in detecting CAD using CCTA and CT-FFR as reference standards. We also analyzed the correlation between CDG values and CT-FFR.

| Study design and population
This retrospective study included 456 patients with suspected CAD who underwent CCTA at Shandong Provincial Hospital from January 2020 to 2022. All patients underwent standard 12-lead ECG examination immediately after a CCTA scan. Patients were excluded if they had known CAD (including previous myocardial infarction, coronary artery bypass grafting, and percutaneous coronary stent implantation), structural heart disease (including myocardial diseases, valvular pathology more than mild, congenital heart disease), atrial fibrillation, heart failure, infective endocarditis, pacemaker implantation, radiofrequency ablation, or poor image quality ( Figure 1). The reference standard for CAD positivity was stenosis on CCTA ≥ 50% or CT-FFR ≤ 0.8. 16,17 All parameters in this study were performed blinded to each other and clinical outcomes, and did not influence clinical treatment. The study was approved by the Ethics Committee of Shandong Provincial Hospital, and the requirement for informed consent was waived.

| CCTA image acquisition
Third-generation dual-source CT (Siemens SOMATOM Force; Siemens Healthcare) was used for image acquisition. Before scanning, patients (except for patients with hypotension) were treated with a sublingually applied nitroglycerine pump spray to dilate the coronary arteries and received breath-hold training to reduce respiratory motion artifacts.  The results of CCTA are diagnosed by two specialists with 10 years of experience in diagnostic cardiovascular imaging, respectively. When there is a disagreement, the two specialists discuss the decision.

| CT-FFR values acquisition
CT-FFR analysis was performed using cFFR software (version 3.2.5; Siemens Healthcare). This software is based on a deep learning model and predicts the FFR values of coronary arteries. After importing the CCTA images into the software, the coronary centerline and lumen were automatically identified and later manually corrected if necessary. Then, a coronary tree was generated with different colors representing different CT-FFR values. The CT-FFR was measured for all vessels of diameter ≥1.8 mm in the coronary tree, the lesionspecific CT-FFR (defined as the per-patient lowest CT-FFR value 2 cm distal to lesion) was recorded in patients with vascular stenosis, and the vessel-specific CT-FFR (defined as the per-patient lowest CT-FFR value in the distal part of vessels) was recorded in patients without plaques. 16 Occluded vessels were assigned a value of 0.5. 19 CT-FFR results were defined as positive if the CT-FFR value ≤0.8. 18

| Invasive coronary angiography (ICA)
Transracial artery puncture was chosen for the ICA procedure.

| Standard 12-lead ECG
The digital 12-lead ECG information of the patients was recorded years of experience in ECG diagnosis, respectively. When there is a disagreement, the two specialists discuss the decision.

| CDG and feature extraction
The CDG was generated as follows. First, the digital 12-lead ECG were filtered to eliminate interference such as power frequency (50 Hz), baseline drift, motion artifact, EMG, and so on. Then, we transformed the 12-lead ECG signals into the 3-lead ECG vector signals by linear transformation matrix of Kors. 20 Based on the lead signal of R wave as the main wave in the vector ECG signals, the R wave peak, J point, T wave apex and T wave endpoint were located, and the ST-T loops in the vector ECG signals were obtained by intercepting the ST-T segments representing the ventricular repolarization process. 21 After that, the ST-T loop of the vectorcardiogram was modeled using a Radial Basis Function neural network to obtain the dynamic ECG information. The CDG was generated by plotting the extracted cardiodynamics information as a 3D graph. The shapes of the CDG represent the dispersion of cardiac repolarization, which is closely related to the degree of myocardial ischemia. 15 These shapes have been shown to be remarkably different between WANG ET AL. | 641 ischemia patients and healthy controls, with the latter presenting with a noticeably regular or annular shape, while CAD patients demonstrate an irregular or nonannular shape.
Then, the shapes of the CDG were interpreted by evaluating the spatial characteristics based on the Lyapunov index and the temporal characteristics based on the Fourier transform of the CDG. 22 The spatial heterogeneity index (SHI) and temporal heterogeneity index (THI) were defined as follows: where N represents the number of data points in the CDG, d n1 represents the distance between the nth data point and its nearest data point, d n2 represents the distance between the nth data point and its nearest data point after 10 steps, and F represents the Fourier transform of the CDG. Finally, a linear support vector machine was used as a classifier to train the screening model for CAD. The CDG value is defined as the distance between the SHI and THI and the classification boundary 23 and was calculated as follows: value CDG value is calculated automatically using the computer program by data analyst who is blind to clinical information.
According to the data of patients with normal or roughly normal chest pain from previous studies, a linear classification boundary was obtained from a CAD detection model established by using machine learning algorithm of linear kernel SVM as follows: The distance of the case samples to the linear classification boundary was a CDG value, the judgment standards were a CDG value ≥0 for positive and <0 for negative.

| CDG value analysis
We found that the positive CAD group had a significantly higher CDG  Figure 2C). Additionally, the CDG value and CT-FFR were correlated (r = −.395; p < .001; Figure 2D). Supporting Information: disadvantages. First, the examination fee is much higher than that of ECG. Second, patients are at risk of radiation exposure, which is forbidden for pregnant women. 26 Although CT-FFR can provide hemodynamic information within the vessels, it is time consuming due to the need to manually adjust the centerline and boundary of the lumen and calculate multivessel fractions. 27 The ECG-derived CDG can also provide functional information, but the calculation time from deterministic learning is far shorter than that of CT-FFR. In our study, we discovered that the ECG-based CDG performed very well in diagnosing suspected CAD populations. This method has potential value in supporting physicians' decisions in the clinical setting.
Diagnosing with the standard 12-lead ECG mainly depends on the physician's visual assessments, resulting in a certain degree of subjectivity, and tiny changes are frequently missed. 11,28 It has been shown that subtle changes in the ST-T segment of the ECG signal can reflect cardiac function and myocardial injury. 29,30 The CDG, based on deterministic learning theory, focuses on analyzing these subtle abnormal signals and detecting the dynamics of cardiac repolarization. 31   Several limitations of this study should be noted. First, this is a single-center retrospective study, and the diagnostic performance of the CDG needs to be confirmed further in a multicenter study with a larger sample size. Second, most of the included population did not meet the criteria for ICA testing, and so ICA or invasive FFR were lacking as a gold standard for validating the diagnostic performance of the CDG, even though many studies have shown that the CT-FFR has equal diagnostic performance to the invasive FFR.

| CONCLUSIONS
The ECG-based CDG generated by deterministic learning can efficiently distinguish CAD and non-CAD. Additionally, this performance was verified with different reference standards and subgroup analyses. This method can be used as a rapid and reliable noninvasive screening tool for suspected CAD patients before the CCTA examination, prompting potential positive patients to undergo further testing early or to reduce unnecessary CCTA testing.

AUTHOR CONTRIBUTIONS
Ying Wang carried out the study, participated in the study design, and