Marzoog BA, Chomakhidze P, Gognieva D, Silantyev A, Suvorov A, Abdullaev M, Mozzhukhina N, Filippova DA, Kostin SV, Kolpashnikova M, Ershova N, Ushakov N, Mesitskaya D, Kopylov P. Development and validation of a machine learning model for diagnosis of ischemic heart disease using single-lead electrocardiogram parameters. World J Cardiol 2025; 17(4): 104396 [DOI: 10.4330/wjc.v17.i4.104396]
Corresponding Author of This Article
Basheer Abdualah Marzoog, MD, Associate Chief Physician, Department of Cardiology, Sechenov University, 8-2 Trubetskaya Street, Moscow 119991, Moskva, Russia. marzug@mail.ru
Research Domain of This Article
Cardiac & Cardiovascular Systems
Article-Type of This Article
Clinical Trials Study
Open-Access Policy of This Article
This article is an open-access article which was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/
Basheer Abdualah Marzoog, Peter Chomakhidze, Daria Gognieva, Artemiy Silantyev, Alexander Suvorov, Magomed Abdullaev, Philipp Kopylov, World-Class Research Center Digital Biodesign and Personalized Healthcare, I.M. Sechenov First Moscow State Medical University (Sechenov University), 8-2 Trubetskaya Street, 119991, Moscow, Russia
Natalia Mozzhukhina, Dinara Mesitskaya, University Clinical Hospital Number 1, Cardiology Department, I.M. Sechenov First Moscow State Medical University (Sechenov University), 8-2 Trubetskaya Street, 119991, Moscow, Russia
Darya Alexandrovna Filippova, Maria Kolpashnikova, Natalya Ershova, Nikolay Ushakov, Undergraduate Medical School student, Sechenov University, Moscow 119991, Moskva, Russia
Sergey Vladimirovich Kostin, Department of Plastic Surgery, Mordovia State University, Saransk 430005, Mordoviya, Russia
Author contributions: Marzoog BA contributed to writing, collecting and analyzing data, interpreting the results, and revising the final version of the manuscript; Suvorov A contributed to biostatistical analysis of the sample; Chomakhidze P, Gognieva D, Silantyev A, Suvorov A, Abdullaev M, Mozzhukhina N, Filippova DA, Kostin SV, Kolpashnikova M, Ershova N, Ushakov N, and Mesitskaya D revised the paper; Kopylov P revised the final version of the manuscript and contributed to the concept and development of the study; All authors read and approved the manuscript.
Supported by Government Assignment, No. 1023022600020-6; RSF Grant, No. 24-15-00549; Ministry of Science and Higher Education of the Russian Federation within the Framework of State Support for the Creation and Development of World-Class Research Center, No. 075-15-2022-304.
Institutional review board statement: The study was approved by the Sechenov University, Russia, from “Ethics Committee Requirement No. 19-23 from 26.10.2023.”
Clinical trial registration statement: The study was registered on clinicaltrials.gov (NCT06181799).
Informed consent statement: All study participants or their legal guardian provided informed written consent about personal and medical data collection prior to study enrolment.
Conflict-of-interest statement: All authors have no conflicts of interest related to the manuscript.
CONSORT 2010 statement: The authors have read the CONSORT 2010 Statement, and the manuscript was prepared and revised according to the CONSORT 2010 Statement.
Data sharing statement: Not applicable.
Open Access: This article is an open-access article that was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution NonCommercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: https://creativecommons.org/Licenses/by-nc/4.0/
Received: December 19, 2024 Revised: February 19, 2025 Accepted: March 31, 2025 Published online: April 26, 2025 Processing time: 123 Days and 14.8 Hours
Abstract
BACKGROUND
Ischemic heart disease (IHD) impacts the quality of life and has the highest mortality rate of cardiovascular diseases globally.
AIM
To compare variations in the parameters of the single-lead electrocardiogram (ECG) during resting conditions and physical exertion in individuals diagnosed with IHD and those without the condition using vasodilator-induced stress computed tomography (CT) myocardial perfusion imaging as the diagnostic reference standard.
METHODS
This single center observational study included 80 participants. The participants were aged ≥ 40 years and given an informed written consent to participate in the study. Both groups, G1 (n = 31) with and G2 (n = 49) without post stress induced myocardial perfusion defect, passed cardiologist consultation, anthropometric measurements, blood pressure and pulse rate measurement, echocardiography, cardio-ankle vascular index, bicycle ergometry, recording 3-min single-lead ECG (Cardio-Qvark) before and just after bicycle ergometry followed by performing CT myocardial perfusion. The LASSO regression with nested cross-validation was used to find the association between Cardio-Qvark parameters and the existence of the perfusion defect. Statistical processing was performed with the R programming language v4.2, Python v.3.10 [^R], and Statistica 12 program.
RESULTS
Bicycle ergometry yielded an area under the receiver operating characteristic curve of 50.7% [95% confidence interval (CI): 0.388-0.625], specificity of 53.1% (95%CI: 0.392-0.673), and sensitivity of 48.4% (95%CI: 0.306-0.657). In contrast, the Cardio-Qvark test performed notably better with an area under the receiver operating characteristic curve of 67% (95%CI: 0.530-0.801), specificity of 75.5% (95%CI: 0.628-0.88), and sensitivity of 51.6% (95%CI: 0.333-0.695).
CONCLUSION
The single-lead ECG has a relatively higher diagnostic accuracy compared with bicycle ergometry by using machine learning models, but the difference was not statistically significant. However, further investigations are required to uncover the hidden capabilities of single-lead ECG in IHD diagnosis.
Core Tip: Ischemic heart disease (IHD) remains the leading cause of mortality and disability globally. The diagnostic methods, including the physical stress test, have poor accuracy (50%) and specificity. The current paper demonstrated that the machine learning model using the parameters of the single-lead electrocardiogram had a diagnostic accuracy of 67% in IHD. Single-lead electrocardiogram has the potential to diagnose IHD using machine learning models.
Citation: Marzoog BA, Chomakhidze P, Gognieva D, Silantyev A, Suvorov A, Abdullaev M, Mozzhukhina N, Filippova DA, Kostin SV, Kolpashnikova M, Ershova N, Ushakov N, Mesitskaya D, Kopylov P. Development and validation of a machine learning model for diagnosis of ischemic heart disease using single-lead electrocardiogram parameters. World J Cardiol 2025; 17(4): 104396
Ischemic heart disease (IHD) remains the leading challenge in terms of mortality and morbidity despite the advances in diagnosis and prevention. However, the early prevention in terms of evaluation of IHD in the early period is still underestimated. Currently, scientists focus more on prevention than on diagnosis and treatment. In this manner, the scientific community developed several cost-effective methods to be confirmed for clinical use for early prevention of IHD, including the use of the single-lead electrocardiography (ECG) and exhaled breath analysis in coronary heart disease prevention[1-4].
IHD diagnosis using single-lead ECG remains in the development stage and requires further elaboration concerning its sensitivity and specificity. Several kinds of single-lead ECG have been used in clinical trials and in the market including Cardio-Qvark[5], Apple Watch[6], Kardia[7], Zio[8], BioHarness[9], Bittium Faros[10], and Carnation Ambulatory Monitor[11,12]. Single-lead ECG has been used to diagnosis myocardial infarction and monitor patients with chronic heart disease and heart failure as well as to classify heartbeat[13-15].
Currently there are several kinds of single-lead ECGs used for commercial purposes and clinical trials. The accuracy and quality of these single-lead ECGs varies[16-18]. The uses of single-lead ECGs include distant monitoring of patients with arrythmias[19-23], such as Holter monitoring[24,25], and monitoring for chronic heart failure[26].
Single-lead ECG has key features that can aid in diagnosing IHD including detecting ischemia through ECG alterations, hemodynamic changes, and clinical signs and symptoms[27]. Additionally, vectorcardiography, a technique that records cardiac electrical activity as closed loops, can be useful for training in ECG and detecting cardiac ischemia[27,28]. Portable and fast electrode placement devices allow for good quality ECG tracings, making single-lead ECG accessible and efficient[27].
Multi-lead ECG in detecting IHD showed that modern ECG systems with vector-based ECG can improve the detection of ECG alterations typical for ischemia compared with the conventional 12-lead ECG[29]. Single-lead consumer ECG devices, such as smartwatches, can be useful for detecting and monitoring arrhythmias but have limitations in detecting ST-segment deviations indicating myocardial infarction or ischemic episodes[30].
The usage of single-lead ECG in IHD [confirmed by computed tomography (CT) myocardial perfusion (CTP) with stress test] has not been previously investigated and requires further elucidation.
MATERIALS AND METHODS
Study design
A prospective, non-randomized, minimally invasive, single-center, case-control cohort study included patients (male and female) aged ≥ 40 years because the risk of coronary heart disease increases dramatically over the age of forty. The recruitment of participants took place from October 27, 2023 to October 28, 2024 at the University Clinical Hospital No 1 of Sechenov University. Initial data of patients with pathology were obtained from the Department of Cardiology and was retrospectively collected from the patient for the healthy patients. The study was conducted in accordance with the standards of Good Clinical Practice and the principles of the Declaration of Helsinki. The study was registered on clinicaltrials.gov (NCT06181799), and the study was approved by the ethical commitment of the Sechenov University, Russia, from “Ethics Committee Requirement No. 19-23 from 26.10.2023.” Informed written consent was collected from the participants to participate in the study and publish the study results and/or any associated figures.
According to the results of the CTP, the participants (n = 80) were divided into two groups. The first group of participants had a stress-induced myocardial perfusion defect (n = 31), and the second group was without a stress-induced myocardial perfusion defect (n = 49) on the CTP. The sample size was reached after calculating the related mean sample power analysis and Pearson correlation power analysis using the SPSS program (Tables 1 and 2).
Failure of the stress test for reasons unrelated to heart disease
Participants with intact mental and physical activity
Diabetes mellitus
Reluctance to continue participating in the study
Written consent to participate in the study, take blood samples, and anonymously publish the results of the study
Presence of signs of acute coronary syndrome (myocardial infarction in the prior 2 days), history of myocardial infarction
Participants in the experimental group are individuals with coronary artery disease, confirmed by stress-induced myocardial perfusion defect on the adenosine triphosphate stress myocardial perfusion computed tomography
Active infectious and non-infectious inflammatory diseases in the exacerbation phase
Cardiac arrhythmias that do not allow exercise ECG testing (Wolff-Parkinson-White syndrome, Sick sinus syndrome, AV block of II-III-degree, persistent ventricular tachycardia)
Diseases of the musculoskeletal system that prevent passing a stress test (bicycle ergometry)
Allergic reaction to iodine and/or adenosine triphosphate
Instrumental tests
Vessel stiffness measurement: Both participant groups underwent comprehensive vascular evaluations using the Fukuda Denshi VaSera VS-1500 system (Japan) to measure arterial stiffness, pulse wave patterns, and biological vascular age. Appropriately sized cuffs were applied to the arms and ankles to assess parameters such as the ABI, arterial rigidity, and estimated vascular age. Electrodes were positioned on both upper limbs, while a microphone for recording cardio-phonograms was securely attached with adhesive tape over the second intercostal space of the sternum. The CAVI, a marker reflecting stiffness of the aorta, femoral, and tibial arteries, was analyzed. Notably, CAVI values are independent of blood pressure fluctuations. They were deemed valid only when consistent across three consecutive heartbeats as standardized in prior research[31]. These measurements were aimed at screening for non-coronary vascular abnormalities and for evaluating the functional age of the vascular system, providing insights into the health of arteries outside the coronary circulation.
Bicycle ergometry: Participants subsequently completed bicycle ergometry assessments using the SCHILLER CS200 system and adhering to either the standard Bruce protocol or its modified variant to evaluate physiological adaptations to exercise. Functional classification of angina (FC) in individuals with abnormal stress test outcomes was stratified using achieved METs and wattage thresholds: FC-III (METs < 4; watts < 50); FC-II (METs 4–7; watts 50–100); and FC-I (METs > 7; watts > 100).
Continuous 12-lead ECG monitoring and manual blood pressure assessments were performed at 2-min intervals during the test, with measurements intensified toward the end of each exercise phase. Termination criteria included: systolic blood pressure exceeding 220 mmHg; ECG-detected horizontal/downsloping ST-segment depression ≥ 1 mm; exercise-induced anginal symptoms; occurrence of ventricular tachycardia or atrial fibrillation; other clinically critical arrhythmias; or attainment of the target heart rate [≥ 86% of the age-adjusted maximum (calculated as 220 minus age)].
Stress CT myocardial perfusion imaging with vasodilatation test using ATP: Prior to stress CTP imaging, all participants underwent renal function assessment through venous creatinine measurement and estimated glomerular filtration rate calculation using the 2021 CKD-EPI creatinine equation (threshold > 30 mL/min/1.73 m²) in accordance with guidelines established by the National Kidney Foundation and American Society of Nephrology[32-35]. Intravenous access was established via the basilic or radial vein for administration of contrast agents and sodium ATP (10 mg/mL) to induce pharmacological stress through increased cardiac workload. The catheter remained in place throughout the procedure for contrast delivery during imaging.
A standardized ATP solution was prepared by diluting 3 mL (30 mg) of ATP with 17 mL of 0.9% sodium chloride resulting in a 20 mL mixture. The injection volume was weight-dependent, calculated at 300 μg/kg over 2 min with specific doses of the diluted solution tailored to individual body mass: 60 kg patients received 12 mL; 70 kg patients 14 mL; 80 kg patients 16 mL; and 100 kg patients 20 mL.
Imaging was performed using a Canon Aquilion Precision 640-slice CT scanner with 0.5 mm slice thickness. The protocol began with non-contrast imaging to evaluate valvular and ascending aortic calcification. Subsequently, 50 mL of Omnipaque contrast agent was administered via catheter to obtain resting myocardial perfusion images. Participants remained supine for 20 min before ATP infusion, which was delivered over 2 min to induce pharmacological stress. Post-stress myocardial perfusion imaging was completed within 30 s of ATP administration using rapid acquisition protocols.
Continuous monitoring ensured adherence to safety and procedural standards. The methodology prioritized precise timing for stress induction and image capture and minimized variability in perfusion measurements. Technical specifications, including slice thickness and contrast volume, were standardized across participants to maintain consistency in data acquisition. This approach reliably compared coronary perfusion dynamics between resting and pharmacologically stressed states while accounting for individual physiological differences through weight-adjusted dosing (Figure 1).
Figure 1 Presentation of the study.
Single-lead electrocardiography using Cardio-Qvark is performed at rest. Subsequently, participants perform exercise via bicycle ergometry (on SCHILLER device c 200; Bruce protocol or modified Bruce protocol). Immediately after bicycle ergometry, Cardio-Qvark is performed again while sitting. CTP: Computed tomography myocardial perfusion; ECG: Electrocardiography; ROC: Receiver operating characteristic.
Cardio-Qvark: All participants at rest passed registration of single-lead ECG and pulse wave before (during 3 min) and just after (during 3 min) the physical stress test (bicycle ergometry) using a portable single-lead recorder (Cardio-Qvark; Russia, Moscow)[36]. The single-lead electrocardiogram and pulse wave results were interpreted using machine learning models developed by the Sechenov University team[36,37]. The Cardio-Qvark parameters were described in detail in the following study and the clinical trial protocol (NCT06181799)[38].
Statistical analysis
For quantitative parameters, distribution normality was assessed using the Shapiro-Wilk test. Descriptive statistics included mean ± SD, median, interquartile range, and minimum/maximum values. Categorical/qualitative features were summarized as proportions (%) and absolute frequencies (n).
For quantitative features, comparative analyses were conducted using Welch’s t-test for normally distributed data (two-group comparisons) and the Mann-Whitney U test for non-normally distributed data (two-group comparisons). For categorical/qualitative features, comparisons were performed with Pearson’s χ2 test; Fisher’s exact test was applied when χ2 assumptions were not met.
For single-lead ECG values, pre-load values (prefixed with “_”) were used, and deltas between immediately after exertion and after 2nd single-lead ECG record were calculated. Statistical processing was carried out using the R programming language v4.2, Python v.3.10 [^R], Statistica 12 program [StatSoft, Inc. (2014). STATISTICA (data analysis software system), version 12. www.statsoft.com.] and SPSS IBM version 28. P values were considered statistically significant at < 0.05.
Outcome determination and predictive feature optimization with cross-validated machine learning framework
Given the modest cohort size (n = 80), a robust resampling strategy was implemented to evaluate predictor reliability. For 1000 iterations, two-thirds of the dataset were randomly selected to identify predictive variables, while the remaining one-third served as a provisional validation subset. This repeated subsampling approach aimed to assess the stability of feature selection and mitigate overfitting risks inherent in smaller datasets[39]. At every preprocessing cycle, quantitative variables underwent normalization and missing value imputation via Bayesian ridge regression. The analysis focused solely on continuous predictors, excluding categorical or binary variables. A gradient boosting classifier was iteratively trained across 1000 cycles to compute feature importance scores at each repetition. Following this, median importance values were derived for each predictor, enabling ranked prioritization of features based on descending median scores.
A refined pipeline incorporated the top ten predictors, which underwent identical preprocessing steps (normalization and Bayesian ridge regression-based imputation). A gradient boosting classifier was subsequently trained, and model performance was evaluated via leave-one-out cross-validation. Post-training, the area under the receiver operating characteristic curve (AUC) was computed, followed by optimization of the classification threshold to derive sensitivity, specificity, and positive/negative predictive values. The AUC of the model was statistically compared to the stress test results using the McNemar test. This workflow was applied independently to Cardio-Qvark-derived data.
RESULTS
The study initially included 101 individuals. After excluding 21 participants (due to voluntary withdrawal or meeting predefined exclusion criteria), the final sample comprised 80 individuals. Baseline characteristics of the cohort are presented in for both the entire sample and subgroups, ensuring comprehensive representation. Continuous variables (e.g., age, clinical measurements) are summarized in Tables 4 and 5.
Table 4 Features of the continuous variables of the sample represented in the table.
Variable
Mean (Min; Max)
SD
Age (year)
56.28 (40.24; 77.94)
10.601
Pulse rest (beat/minute)
70.29 (49.00; 93.00)
9.559
SBP rest (mmHg)
123.16 (54.00; 159.00)
15.437
DBP rest (mmHg)
80.61 (60.00; 122.00)
11.238
Body weight (kg)
77.92 (52.50; 140.00)
16.236
Height (cm)
169.95 (148.00; 190.00)
8.835
BMI (kg/m2)
26.93 (18.49; 48.44)
4.901
Pulse rest (beat/minute)
69.68 (49.00; 99.00)
9.260
Pulse after stress (beat/minute)
86.89 (63.00; 115.00)
10.579
Goal heart rate (beat/minute)
163.72 (142.06; 179.76)
10.601
Max HR
146.25 (108.00; 199.00)
14.111
Reached %
89.53 (64.89; 135.25)
9.166
WT
125.63 (75.00; 250.00)
44.111
METs
6.67 (2.90; 11.90)
1.977
EF (%)
85.28 (60.50; 104.50)
8.358
Vessel age (years)
56.50 (20.00; 80.00)
13.520
R-CAVI
8.21 (4.80; 15.10)
1.379
L-CAVI
8.18 (4.80; 14.90)
1.299
Mean CAVI
8,194 (4,800; 15,000)
1,331
RABI
1.15 (0.86; 1.40)
0.088
LABI
1.15 (0.89; 1.42)
0.084
Mean SBP B (mmHg)
134.38 (105.50; 169.00)
13.086
Mean DBP B (mmHg)
85.28 (60.50; 104.50)
8.358
BP RB [(SBP + DBP)/2] (mmHg)
103.98 (75.00; 137.00)
11.547
BP LB [(SBP + DBP)/2] (mmHg)
104.54 (71.00; 136.00)
10.534
Mean BP B (mmHg)
104.26 (73.00; 136.50)
10.772
BP RA [(SBP + DBP)/2] (mmHg)
108.39 (80.00; 137.00)
12.553
BP LA [(SBP + DBP)/2] (mmHg)
108.63 (81.00; 138.00)
11.519
Mean BP A (mmHg)
108.51 (81.50; 135.00)
11.468
Mean ABI
1.15 (0.88; 1.41)
0.081
RTb
80.53 (59.00; 152.00)
13.441
LTb
77.26 (58.00; 128.00)
14.012
Mean Tb
78.89 (59.00; 132.50)
12.547
Right Tba
86.26 (23.00; 117.00)
16.232
Left Tba
86.40 (24.00; 114.00)
14.685
Mean Tba
86.33 (23.50; 115.00)
15.358
Lha (cm)
148.17 (130.33; 164.47)
7.182
haPWV (m/s)
0.91 (0.67; 1.42)
0.124
β-stiffness index from PWV
2.83 (1.21; 7.04)
0.813
Creatinine (µmol/L)
82.74 (53.90; 138.00)
16.014
eGFR (2021 CKD-EPI creatinine)
85.31 (45.40; 113.70)
14.684
Table 5 Features of the categorical variables of the participants.
Index
Factor
Absolute value, relative value
Gender
F
39 (48.75)
M
41 (51.25)
Obesity stage
Normal
30 (37.500)
Overweight
29 (36.25)
1 degree
20 (25.00)
3 degree
1 (1.25)
Smoking
Yes
14 (17.50)
No
66 (82.50)
Concomitant diseases
Yes
41 (51.25)
No
35 (43.75)
Missing data
4 (5.00)
Atherosclerosis of the coronary artery
Yes
31 (38.75)
No
49 (61.25)
Hemodynamically significant coronary artery atherosclerosis on the CTP (> 60% stenosis)
Yes
9 (11.25)
No
71 (88.75)
Stress-induced myocardial perfusion defect on the CTP
Yes
31 (38.75)
No
49 (61.25)
Myocardial perfusion defect before stress ATP on the CTP
Yes
26 (32.50)
No
54 (67.50)
Atherosclerosis in other arteries
Yes
41 (51.25)
No
32 (40.00)
Missing data
7 (8.75)
Atherosclerotic vascular (namely)
Carotid
1 (1.25)
Carotid brachiocephalic bifurcation
41 (51.25)
Missing data
38 (47.50)
Carotid artery atherosclerosis
Yes
39 (48.75)
No
34 (42.50)
Missing data
7 (8.750)
Brachiocephalic artery atherosclerosis
Yes
37 (46.25)
No
36 (45.00)
Missing data
7 (8.75)
Arterial hypertension
Yes
40 (50.00)
No
40 (50.00)
Stage of the arterial hypertension
I
5 (6.25)
II
20 (25.00)
III
16 (20.00)
Degree of hypertension
Degree 1
19 (23.75)
Degree 2
13 (16.25)
Degree 3
9 (11.25)
Risk of cardiovascular disease
Low
27 (33.75)
Moderate
27 (33.75)
High
18 (22.50)
Very high
8 (10.00)
SCAD from anamnesis
Yes
3 (3.75)
No
29 (36.25)
Missing data
48 (60.00)
Blood pressure reaction type on stress test
Asthenic
4 (5.00)
Hypotonic
4 (5.00)
Hypertonic
8 (10.00)
Normotonic
64 (80.00)
Functional class by Watt
FC-I
8 (10.00)
FC-II
9 (11.25)
No SCAD
63 (78.75)
Functional class by METs
FC-I
6 (7.50)
FC-II
10 (12.50)
FC-III
1 (1.25)
No SCAD
63 (78.75)
Reaction type to stress test (positive/negative)
Negative
42 (52.50)
Suspected
21 (22.5)
Positive
17 (21.25)
Reason for discontinuation of the stress test
Horizontal ST depression > 1 mm
8 (10)
Reach goal HR
72 (90)
Tolerance to exertion on stress test
Low
2 (2.50)
Moderate
43 (53.75)
Close to high
8 (10.00)
High
16 (20.00)
Very high
11 (13.75)
CAVI degree
Normal (< 8)
36 (45.00)
Borderline (8-9)
22 (27.50)
Pathological (> 9)
22 (27.50)
ABI degree
Normal
76 (95.00)
Borderline
2 (2.50)
Abnormal
1 (1.25)
Noncompressible
1 (1.25)
Biological estimated vascular age
Normal
45 (56.25)
High
35 (43.75)
CKD stage
I
35 (43.75)
II
41 (51.25)
IIIa
4 (5.00)
The cohort was stratified into subgroups based on the presence or absence of ATP-induced myocardial perfusion defects identified via CTP imaging. Comparative demographic and clinical characteristics of these subgroups are detailed in Tables 6 and 7.
The diagnostic accuracy of the physical stress test
We evaluated the diagnostic performance of standard bicycle ergometry by analyzing receiver operating characteristic curves. The predictor variable was defined as a positive exercise-induced response (Reaction type = Positive), and the outcome variable was the presence of myocardial perfusion defects post-ATP stress (Myocardial_perfusion_defect_ after_stress_ATP). Key performance metrics (e.g., AUC, sensitivity, specificity) are summarized in Table 8.
Two machine learning methodologies, LASSO logistic regression and XGBoost, were applied to identify Cardio-Qvark parameters predictive of IHD. Prior to analysis, all Cardio-Qvark parameters were standardized to ensure comparability. A 5-fold cross-validation framework was employed for simultaneous model training and feature selection, with no additional dataset partitioning due to the limited sample size (n = 80), ensuring retention of statistical power.
In LASSO regression, features were selected based on the absolute values of regression coefficients, leveraging L1 regularization to eliminate non-contributory variables. XGBoost in contrast utilized gain scores to quantify feature importance, reflecting their cumulative contribution to predictive accuracy across decision trees[40]. The feature selection process aimed to retain the minimal subset of predictors achieving an AUC threshold of ≥ 0.6703 during preliminary validation. Model performance was evaluated using AUC, sensitivity, and specificity metrics, with the top ten predictors ranked by median importance scores detailed in Table 9.
The model was iteratively refined by reconstructing it with the five highest-impact predictors from Table 8 as determined by median importance rankings. A leave-one-out cross-validation procedure was then implemented, standardizing quantitative predictors at each iteration to minimize scaling biases. This process generated estimates of sensitivity, specificity, positive predictive value, and negative predictive value, with results summarized in Table 10.
Table 10 Single-lead electrocardiography (Cardio-Qvark) in the diagnosis of ischemic heart disease.
Chars
Point estimate
95%CI
AUC
0.670
0.531-0.802
Sensitivity
0.516
0.333-0.695
Specificity
0.755
0.628-0.88
Negative predictive value
0.712
0.586-0.83
Positive predictive value
0.571
0.387-0.758
Bootstrap resampling (1000 replicates) derived confidence intervals for reliability assessment, while the McNemar test facilitated a comparative analysis against stress-test outcomes[41] (Figure 2).
Figure 2 Diagnostic accuracy of the load test and the built model.
There was no statistical significance, but the model had better predictive properties, P = 0.337. AUC: Area under the receiver operating characteristic curve.
DISCUSSION
Using single-lead ECG in the diagnosis of IHD is a potential novel diagnostic strategy that requires further elaboration. Moreover, the usage of single-lead ECG in optimizing the physical stress test such as bicycle ergometry is an optimistic strategy[1]. Single-lead ECG results can be interpreted using machine learning models to increase the diagnostic accuracy and reduce the time needed for the interpretation of the results by physicians. Additionally, the primary data of the single-lead ECG (approximately 200 parameters) can be used as a novel risk scoring for future IHD development or death[14]. The current paper showed a 67% diagnostic accuracy for the Cardio-Qvark in the diagnosis of IHD and did not reach the cutoff where statistical significance can be achieved. When the AUC values are higher, a better performance is achieved. Moreover, a good negative predictive value suggests a good ability of the model to exclude patients without IHD. However, a low positive predictive value suggests a poor ability of the model to detect IHD. Therefore, the model is good in terms of avoiding the true negative patients.
Using the physical stress test monitored by 12-lead ECG remains the elementary test for the primary detection of IHD. However, severe limitations exist in the diagnostic accuracy related to the ECG artifacts during the movement of patients during the physical stress test. This issue can be overcome by combining the single-lead ECG with the physical stress test and performing it immediately after the physical stress test. The suggested parameters of our model include q0_QTc, q0_Beta, q0_J80A, q0_Pan…30, and q0_SA.
Improving the diagnostic accuracy of the physical stress test is a point of focus of the cardiological scientific community. Several attempts were performed to enhance the diagnostic performance of the physical exertion tests using complementary methods such as the dynamics of cardiac electrical activity during exercise testing[42]. The study suggests that incorporating the equivalent electric cardiac generator of dipole type during exercise ECG testing can enhance the accuracy of diagnosing coronary artery disease[43].
A previous clinical study using a wearable wireless ECG has shown that the single-lead ECG has a poor sensitivity of 8.3% (1.0%-27.0%) and high specificity 89.9% (80.2%-95.8%) for detection of reversible IHD[44]. The study concluded that both 12-lead ECG [sensitivity 12.5% (3.0%-34.4%) and specificity 91.3% (82.0%-96.7%)] and the single-lead ECG have poor clinical usefulness in terms of the ability to detect IHD. Interestingly, a dramatic difference has been observed in the II lead of the 12-lead ECG compared with the single-lead ECG[44]. However, another study suggested that the use of deep learning models could enhance the diagnostic accuracy (sensitivity) of the ECG for IHD in the emergency department[45].
Advancements in single-lead ECG technology for the detection of IHD demonstrated that machine learning models based on single-lead ECG and pulse wave parameters, along with age and gender, can simplify screening diagnostics of ejection fraction decrease and diastolic dysfunction with high accuracy[46]. Furthermore, high-frequency ECG signals have shown increased sensitivity and early timing in diagnosing cardiac ischemia, and portable high-resolution ECG devices have demonstrated utility in acute emergency settings[46].
Several clinical trials to assess the reliability of single-lead ECG in the diagnosis of IHD and arrythmia in both adults and children are ongoing (NCT05756309, NCT06181799).
Study limitations included the relatively small sample size and the absence of external validation. However, this methodology is the first of its kind globally. The current results can be reevaluated through the implication of a larger sample size. However, a better performance of the machine learning model has been observed and using a larger sample with external validation will potentially give a better results. The current findings can be a basis for future studies to evaluate the diagnostic accuracy of the single-lead ECG in the diagnosis of IHD. Further investigations are required to evaluate the hidden potential of single-lead ECG in IHD diagnosis.
CONCLUSION
Single-lead ECG (Cardio-Qvark) has the potential to be used as an additional method for in diagnosis of IHD in combination with the physical stress test such as bicycle ergometry. Further clinical studies are required on a larger sample size to validate the usage of the Cardio-Qvark in clinical practice for the diagnosis of IHD. The current work is the first of its kind globally and is considered the starting point for investigating IHD using a confirmed method such as the CTP with stress test. The clinical application of single-lead ECG is not limited to diagnosis. It could be used for prognosis and prevention through the early detection of IHD risk stratification.
Footnotes
Provenance and peer review: Unsolicited article; Externally peer reviewed.
Peer-review model: Single blind
Corresponding Author’s Membership in Professional Societies: European Society of Cardiology, No. 1137915.
Specialty type: Cardiac and cardiovascular systems
Country of origin: Russia
Peer-review report’s classification
Scientific Quality: Grade B, Grade D
Novelty: Grade B, Grade C
Creativity or Innovation: Grade B, Grade C
Scientific Significance: Grade B, Grade B
P-Reviewer: Jiao Y; Xue Y S-Editor: Lin C L-Editor: Filipodia P-Editor: Zhao YQ
Marzoog BA, Chomakhidze P, Gognieva D, Gagarina NV, Silantyev A, Suvorov A, Fominykha E, Mustafina M, Natalya E, Gadzhiakhmedova A, Kopylov P. Machine Learning Model Discriminate Ischemic Heart Disease Using Breathome Analysis.Biomedicines. 2024;12:2814.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Reference Citation Analysis (0)]
Abdou A, Krishnan S. Horizons in Single-Lead ECG Analysis From Devices to Data.Front Signal Process. 2022;2:866047.
[PubMed] [DOI] [Full Text][Cited in This Article: ]
Avram R, Ramsis M, Cristal AD, Nathan V, Zhu L, Kim J, Kuang J, Gao A, Vittinghoff E, Rohdin-Bibby L, Yogi S, Seremet E, Carp V, Badilini F, Pletcher MJ, Marcus GM, Mortara D, Olgin JE. Validation of an algorithm for continuous monitoring of atrial fibrillation using a consumer smartwatch.Heart Rhythm. 2021;18:1482-1490.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 17][Cited by in RCA: 42][Article Influence: 10.5][Reference Citation Analysis (0)]
Perez MV, Mahaffey KW, Hedlin H, Rumsfeld JS, Garcia A, Ferris T, Balasubramanian V, Russo AM, Rajmane A, Cheung L, Hung G, Lee J, Kowey P, Talati N, Nag D, Gummidipundi SE, Beatty A, Hills MT, Desai S, Granger CB, Desai M, Turakhia MP; Apple Heart Study Investigators. Large-Scale Assessment of a Smartwatch to Identify Atrial Fibrillation.N Engl J Med. 2019;381:1909-1917.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 810][Cited by in RCA: 1056][Article Influence: 176.0][Reference Citation Analysis (0)]
Tison GH, Sanchez JM, Ballinger B, Singh A, Olgin JE, Pletcher MJ, Vittinghoff E, Lee ES, Fan SM, Gladstone RA, Mikell C, Sohoni N, Hsieh J, Marcus GM. Passive Detection of Atrial Fibrillation Using a Commercially Available Smartwatch.JAMA Cardiol. 2018;3:409-416.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 262][Cited by in RCA: 280][Article Influence: 46.7][Reference Citation Analysis (0)]
Nault I, André P, Plourde B, Leclerc F, Sarrazin JF, Philippon F, O'Hara G, Molin F, Steinberg C, Roy K, Blier L, Champagne J. Validation of a novel single lead ambulatory ECG monitor - Cardiostat™ - Compared to a standard ECG Holter monitoring.J Electrocardiol. 2019;53:57-63.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 20][Cited by in RCA: 15][Article Influence: 2.5][Reference Citation Analysis (0)]
Zhao X, Zhang J, Gong Y, Xu L, Liu H, Wei S, Wu Y, Cha G, Wei H, Mao J, Xia L. Reliable Detection of Myocardial Ischemia Using Machine Learning Based on Temporal-Spatial Characteristics of Electrocardiogram and Vectorcardiogram.Front Physiol. 2022;13:854191.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited in This Article: ][Cited by in Crossref: 12][Reference Citation Analysis (0)]
Vardas P, Cowie M, Dagres N, Asvestas D, Tzeis S, Vardas EP, Hindricks G, Camm J. The electrocardiogram endeavour: from the Holter single-lead recordings to multilead wearable devices supported by computational machine learning algorithms.Europace. 2020;22:19-23.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 8][Cited by in RCA: 8][Article Influence: 2.0][Reference Citation Analysis (0)]
Attia ZI, Harmon DM, Dugan J, Manka L, Lopez-Jimenez F, Lerman A, Siontis KC, Noseworthy PA, Yao X, Klavetter EW, Halamka JD, Asirvatham SJ, Khan R, Carter RE, Leibovich BC, Friedman PA. Prospective evaluation of smartwatch-enabled detection of left ventricular dysfunction.Nat Med. 2022;28:2497-2503.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 3][Cited by in RCA: 56][Article Influence: 18.7][Reference Citation Analysis (0)]
De Luna AB.
Limitations of the Conventional ECG: Utility of Other Techniques. In: Clinical Electrocardiography: A Textbook, Fourth Edition. New Jersey: Wiley & Sons, Ltd, 2012: 552-570.
[PubMed] [DOI] [Full Text][Cited in This Article: ]
Braun T, Spiliopoulos S, Veltman C, Hergesell V, Passow A, Tenderich G, Borggrefe M, Koerner MM. Detection of myocardial ischemia due to clinically asymptomatic coronary artery stenosis at rest using supervised artificial intelligence-enabled vectorcardiography - A five-fold cross validation of accuracy.J Electrocardiol. 2020;59:100-105.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 9][Cited by in RCA: 12][Article Influence: 2.4][Reference Citation Analysis (0)]
Delgado C, Baweja M, Crews DC, Eneanya ND, Gadegbeku CA, Inker LA, Mendu ML, Miller WG, Moxey-Mims MM, Roberts GV, St Peter WL, Warfield C, Powe NR. A Unifying Approach for GFR Estimation: Recommendations of the NKF-ASN Task Force on Reassessing the Inclusion of Race in Diagnosing Kidney Disease.Am J Kidney Dis. 2022;79:268-288.e1.
[RCA] [PubMed] [DOI] [Full Text][Cited in This Article: ][Cited by in Crossref: 230][Cited by in RCA: 420][Article Influence: 140.0][Reference Citation Analysis (0)]
Gognieva D, Vishnyakova N, Mitina Y, Chomakhidze P, Mesitskaya D, Kuznetsova N, Khiari M, Ryabykina G, Boytsov S, Syrkin A, Saner H, Kopylov P. Remote Screening for Atrial Fibrillation by a Federal Cardiac Monitoring System in Primary Care Patients in Russia: Results from the Prospective Interventional Multicenter FECAS-AFS Study.Glob Heart. 2022;17:4.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited in This Article: ][Reference Citation Analysis (0)]
Cawley GC, Talbot NLC. On Over-fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation.J Mach Learn Res. 2010;11:2079-2107.
[PubMed] [DOI][Cited in This Article: ]
Marius OU, Happiness OI. An extended McNemar test for comparing correlated proportion of positive responses.Biom Biostat Int J. 2019;8:125-137.
[PubMed] [DOI] [Full Text][Cited in This Article: ]
Kramm MN, Strelkov NO, Chomakhidze PS, Kopylov FY. Study of additional diagnostic signs of myocardial ischemia.Kardiol serdechno-sosud khir. 2016;9:52.
[PubMed] [DOI] [Full Text][Cited in This Article: ]
Lee BT, Kwon JM, Cho J, Bae W, Park H, Seo WW, Cho I, Lee Y, Park J, Oh BH, Jeon KH. Usefulness of Deep-Learning Algorithm for Detecting Acute Myocardial Infarction Using Electrocardiogram Alone in Patients With Chest Pain at Emergency Department: DAMI-ECG Study.J Cardiovasc Interv. 2023;2:100-112.
[PubMed] [DOI] [Full Text][Cited in This Article: ]
Kuznetsova N, Sagirova Z, Suvorov A, Dhif I, Gognieva D, Afina B, Poltavskaya M, Sedov V, Chomakhidze P, Kopylov P. A screening method for predicting left ventricular dysfunction based on spectral analysis of a single-channel electrocardiogram using machine learning algorithms.Biomed Signal Process Control. 2023;86:105219.
[PubMed] [DOI] [Full Text][Cited in This Article: ]
Srinivasan A, K V, Padmanabhan D, Prabhavathi, K S, Gopal S.
Early detection of Ischaemia Through High Frequency ECGs: The role of Medical-Grade Wearables for Chest Pain Triages. 2022 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT); 2022.
[PubMed] [DOI] [Full Text][Cited in This Article: ]
