Skip to main content
Download PDF

Screening for severe coronary stenosis in patients with apparently normal electrocardiograms based on deep learning

BMC Medical Informatics and Decision Making volume 24, Article number: 355 (2024) Cite this article

Abstract

Background

Patients with severe coronary arterystenosis may present with apparently normal electrocardiograms (ECGs), making it difficult to detect adverse health conditions during routine screenings or physical examinations. Consequently, these patients might miss the optimal window for treatment.

Methods

We aimed to develop an effective model to distinguish severe coronary stenosis from no or mild coronary stenosis in patients with apparently normal ECGs. A total of 392 patients, including 138 with severe stenosis, were selected for the study. Deep learning (DL) models were trained from scratch and using pre-trained parameters via transfer learning. These models were evaluated based on ECG data alone and in combination with clinical information, including age, sex, hypertension, diabetes, dyslipidemia and smoking status.

Results

We found that DL models trained from scratch using ECG data alone achieved a specificity of 74.6% but exhibited low sensitivity (54.5%), comparable to the performance of logistic regression using clinical data. Adding clinical information to the ECG DL model trained from scratch improved sensitivity (90.9%) but reduced specificity (42.3%). The best performance was achieved by combining clinical information with the ECG transfer learning model, resulting in an area under the receiver operating characteristic curve (AUC) of 0.847, with 84.8% sensitivity and 70.4% specificity.

Conclusions

The findings demonstrate the effectiveness of DL models in identifying severe coronary stenosis in patients with apparently normal ECGs and validate an efficient approach utilizing existing ECG models. By employing transfer learning techniques, we can extract “deep features” that summarize the inherent information of ECGs with relatively low computational expense.

Peer Review reports

Background

Coronary artery disease (CAD) is one of the most common cardiovascular diseases in the world, leading to severe life-threatening cardiovascular events such as acute myocardial infarction (MI), malignant arrhythmia, and heart failure [1, 2]. Timely treatment of CAD can improve the prognosis and reduce the incidence of adverse events and mortality.

Early screening and diagnosis of CAD are thus of great significance. While coronary angiography (CAG) is considered the “gold standard” for diagnosing CAD, it is not suitable for screening due to its invasive nature, high cost, and the need for specialized equipment and personnel [3]. In contrast, electrocardiogram (ECG) is a simple and widely used test for both screening and diagnosing heart diseases [4]. However, ECG has limited sensitivity and specificity in predicting coronary artery stenosis [5], especially in CAD patients without myocardial ischemia, whose ECGs may appear normal or “apparently normal”. Figure 1 illustrates this issue with two patients having apparently normal resting ECGs. Patient A’s CAG images show no obvious coronary lesions, whereas Patient B’s CAG images reveal severe coronary stenosis, with about 95% stenosis in the distal part of the circumflex branch and about 80% stenosis in the middle part of the left anterior descending branch. The stenosis percentages are calculated using the quantitative coronary angiography (QCA) technique, which compares the diameter of the narrowed artery segment to that of a healthy reference segment. In clinical practice, it is common for CAD patients with severe coronary stenosis to present with only mild angina and apparently normal ECGs [6]. This scenario is particularly harmful because such patients’ conditions may not be detected during routine physical examinations or screenings, potentially missing the optimal window for treatment by the time severe symptoms appear.

Fig. 1
figure 1

ECGs and CAG Images of Two Patients. Both patients have an apparently normal ECG. Patient A’s CAG images show no obvious coronary lesions; however, Patient B’s CAG images show severe coronary stenosis in both the circumflex branch and the left anterior descending branch (see red arrows)

Full size image

In recent years, artificial intelligence (AI) have been increasingly applied in cardiovascular studies [7,8,9]. Deep learning (DL) models based on ECG data have been developed for diagnosing arrhythmia [10,11,12], cardiac systolic or diastolic dysfunction [13,14,15], valvular heart disease [16, 17], cardiomyopathy [18, 19], predicting the risk of death [20, 21], and other conditions [22, 23]. For CAD, Leasure’s group developed a DL model that correlates ECGs with CAG results to classify the severity and location of coronary artery stenosis [24]. Similarly, Huang and colleagues tested six well-known computer vision DL models to predict the location of obstructed coronary artery from ECG images [25]. Both group’s models obtained high sensitivity and high specificity. Moreover, a recent study demonstrated the possibility and feasibility of developing a DL model to detect CAD using only ECG data, achieving moderate performance [26]. However, these studies do not emphasize the status of the ECG used and might include those showing obvious abnormalities. Such abnormalities could reduce the AI model’s effectiveness, as doctors can typically identify these issues without AI assistance.

We carefully selected apparently normal ECGs for our study to explore the performance of DL models in identifying patients with severe (≥ 90%) stenosis versus no/mild (< 50%) stenosis, as supported by [27,28,29]. Our findings are as follows:

  1. 1

    DL models have the potential to predict severe coronary stenosis in patients with apparently normal ECGs.

  2. 2

    The DL model trained from scratch using only ECG data can achieve a specificity of 0.746, but its sensitivity is low. This performance is comparable to that of logistic regression using clinical information such as age, gender, hypertension, diabetes, dyslipidemia, and smoking status.

  3. 3

    Adding clinical information to the ECG DL model trained from scratch can improve sensitivity, but it significantly reduces the specificity.

  4. 4

    Combining clinical information with the outputs of an existing ECG classification model (transfer learning) can achieve an accuracy of 0.750, with a sensitivity of 0.848 and specificity of 0.704 for identifying severe coronary artery stenosis.

Methods

Data sources

Patients admitted to the Department of Cardiology at the Second Hospital of Tianjin Medical University from January 2019 to February 2021, who received at least one standard 12-lead ECG within 3 days before CAG, were included in the study. The selection criteria were as follows: over 18 years of age; apparently normal ECG; no acute or old MI; no previous percutaneous coronary intervention (PCI) or coronary artery bypass grafting (CABG). An apparently normal ECG is defined by the following [30]: 1) heart rate between 55 and 100 bpm; 2) presence of P waves in all leads, with normal morphology; 3) presence of a QRS complex following each P wave; 4) PR interval between 0.12 to 0.20 seconds; 5) QRS duration ≤ 0.12 seconds, absence of pathological Q waves, and no ventricular hypertrophy; 6) QTc interval in males ≤ 440 ms or females ≤ 460 ms; 7) QRS axis without deviation; 8) normal increase of R wave in precordial leads; 9) ST segment without elevation or depression; 10) smooth and asymmetric T waves. We also collect data on age, gender, past medical history, smoking status, and results from laboratory testing. All clinical data were obtained with the informed consent of patients and approved by the hospital ethics committee. A summary of the clinical baseline characteristics for the entire dataset is shown in Table 1.

Table 1 Baseline Characteristics for All Data
Full size table

A total of 392 patients were selected for the study, and each patient had one 10-second standard 12-lead ECG acquired using a Fukuda Denshi CardiMax FX-7402 ECG Machine. A professional cardiologist, with over 5 years of experience and qualified for ECG diagnosis, manually interpreted the ECG graphs printed at a paper speed of 25mm/s and a voltage calibration of 10mm/mV. We decrypted and extracted raw ECG data that had a 500Hz sampling frequency using the EFS-200C software provided by the manufacturer. A summary of the ECG measurements is listed in Table 2.

Table 2 ECG Measurements for All Data.(± is for verified normally distributed data, (−) is the interquartile range for non-normally distributed data)
Full size table

Based on the CAG results, we classified coronary artery stenosis as severe (case group) and no/mild (control group) by visual inspection of lumen diameter. Severe stenosis was defined as ≥ 90% stenosis in at least one of the major coronary arteries or their major branches, while no/mild stenosis was defined as < 50% stenosis in all major coronary arteries and their major branches. The CAG results were independently assessed and reported by two professional interventional cardiologists, each with over 10 years of experience, using computer-assisted QCA results as a reference. The final degree of stenosis was determined through their mutual agreement. The lesion information from CAG is shown in Table 3. To simplify the classification process and enhance the model’s robustness in distinguishing between significant (severe) and non-significant (no/mild) stenosis, we combined the no and mild stenosis categories. Under the guidance of cardiologists, technicians labeled the ECGs as follows: 0 for no/mild stenosis and 1 for severe stenosis.

Table 3 Lesion Information from CAG for All Data
Full size table

We randomly divided the data into training and test sets in a ratio of approximately 3:1, resulting in 288 training examples and 104 test examples. Since only one ECG recording was selected per subject, no samples from the same subject appeared in both sets. The flow diagram for the overall data selection process is shown in Figure 2, and Table 4 lists the baseline characteristics for the training and test sets.

Fig. 2
figure 2

Flow diagram of the criteria

Full size image
Table 4 Baseline characteristics for training and test set
Full size table

Models

In the study, we conducted experiments to differentiate the severity of coronary stenosis using ECGs based on a one-dimensional deep convolutional neuron network (1D CNN) structure, which has been verified to perform well on ECG-related tasks [31,32,33]. The input layer of the model structure contained 12 channels representing the 12 leads of the ECG, and each channel had 5000 data points from a 10-second recording with a sampling frequency of 500 Hz. The original output layer had 47 units corresponding to 47 ECG classes in the original classification task. To adapt the model for our task, we added a fully connected layer with one output unit to produce the probability of severe coronary stenosis (Table 5).

Table 5 Model architecture. ECG length n = 5000, number of leads d = 12, and output dimension c = 47. The first dimension ∗ represents number of samples in a batch
Full size table

Our experiments involved four DL models based on different input types and training methods: (1) a model trained from scratch using only ECG data, (2) a model trained from scratch using ECG data and baseline clinical features, (3) a transfer learning model with ECG data alone, and (4) a transfer learning model with ECG and baseline features (first four models in Figure 3). For models trained from scratch, initial weights were set randomly and updated based on binary cross-entropy between the predicted stenosis probabilities and the actual labels. For transfer learning models, we leveraged pre-trained weights from the ECG classification task, which likely encode ECG-specific information, and fine-tuned only the last two dense layers to optimize for our specific task. The initial learning rate for models trained from scratch was set to 1 × 10−6, and decreasing by 5% each epoch, with a dropout rate of 0.2 applied at the end of each convolutional layer. Transfer learning models used a higher learning rate of 1 × 10−3 for the trainable layers, leveraging pre-trained ECG knowledge to accelerate adaptation without destabilization.

Fig. 3
figure 3

Diagram of experimental models

Full size image

Across all four models, a batch size of 32 was used to balance memory efficiency and training stability, with the Adam optimizer applied to accelerate convergence and enhance model performance. Training for each model was set to stop when the validation loss ceased to decrease, as shown in Figure 4. In this figure, subplots A and B illustrate the loss curves for models trained from scratch, while subplots C and D display those for transfer learning models. To provide a comprehensive view of the training process, we recorded and displayed the training and validation losses for each model over the course of up to 1000 epochs.

Fig. 4
figure 4

Training and validation loss over epochs for dl models

Full size image

We refer to the 47 features extracted from the original deep CNN model as “deep features”, and the details of the features are discussed in the Discussion section. For the models with additional baseline features, we added the six baseline characteristics of patients (age (normalized), gender (0 female; 1 male), hypertension (0 no; 1 yes), diabetes (0 no; 1 yes), dyslipidemia (0 no; 1 yes), smoking status (0 no; 1 yes)) on the second-to-last layer. This combination resulted in a total of 53 features, which were fed into the last fully-connected layer to obtain the final decision.

We also conducted logistic regression on the baseline features alone to study the effects of these characteristics on differentiating severe coronary stenosis. Logistic regression was chosen for its ability to provide insights into the relationships between baseline features and the probability of severe stenosis [34]. This approach allowed us to isolate the impact of baseline features and verify the outcomes when combined with deep features. All the five experimental models are illustrated in Figure 3.

Evaluation metrics

After training, we calculated and reported the results metrics on the test data set, which included area under the curve (AUC), accuracy, precision, sensitivity, specificity, and F1-score. The AUC refers specifically to the area under the receiver operating characteristics (ROC) curve. A ROC curve plots the performance of a binary classification model at different threshold settings, and AUC represents the overall classification ability of the model. The x-axis of the ROC curve is the false positive rate (FPR), and the y-axis is the true positive rate (TPR). In our study, a false positive (FP) is an actual no/mild stenosis predicted as severe stenosis by the model, and the true positive (TP) is an actual severe stenosis predicted correctly as severe stenosis. FPR is the ratio of FP to the total actual no/mild stenosis (FP + true negative (TN)), while TPR is the ratio of TP to the total actual severe stenosis (TP + false negative (FN)). We preferred a higher TPR as well as a lower FPR and selected a threshold that maximized the Youden’s Index (TPR - FPR). Based on this selected threshold, we calculated and reported the accuracy, precision, sensitivity, specificity, and F1-score. The equations used for the calculations are as follows:

$$accuracy=\frac{TP\;+\;TN}{TP\;+\;TN\;+\;FP\;+\;FN}$$
(1)
$$precision=\frac{TP}{TP\;+\;FP}$$
(2)
$$sensitivity\;(same\;as\;TPR)=\frac{TP}{TP\;+\;FN}$$
(3)
$$specificity\;(same\;as\;(1-TPR))\;=\frac{TN}{TN\;+\;FP}$$
(4)
$$F1-score=\frac{2\;\times\;precision\;\times\;sensitivity}{precision\;+\;sensitivity}$$
(5)

Results

Using the testing data, we obtained the probabilities of severe coronary stenosis for each model. Figure 5 shows the ROC curves of all the models. The model combining ECG and baseline features with transfer learning achieved the highest AUC (0.847), while other DL models had similar AUC values (0.649 - 0.665). The logistic regression model for baseline features had an AUC of 0.717, the second highest value.

Fig. 5
figure 5

ROC curves of all models

Full size image

Table 6 presents the metrics of the test results for each model, with the optimal threshold selected based on Youden Index. The model combining ECG and baseline features with transfer learning achieved the best accuracy (0.750), precision (0.571) and F1-score (0.683) among the experimental models. The ECG and baseline features model trained from scratch had the highest sensitivity (0.909), while the ECG-only model trained from scratch had the highest specificity (0.746). The AUC, sensitivity, and specificity for each model are shown in Figure 3, and the normalized confusion matrices of the models based on the selected thresholds are shown in Figure 6.

Table 6 Metrics of All Models (±95% CI)
Full size table
Fig. 6
figure 6

Normalized confusion matrices of all models

Full size image

We also conducted regression analysis on the baseline features, and these results are presented in Table 7. Age, hypertension and dyslipdemia had significant coefficients among all the baseline features.

Table 7 Regression analysis results for baseline features
Full size table

Discussion

In the study, we investigated the feasibility of using DL techniques to distinguish between severe (≥ 90%) and no/mild (< 50%) coronary stenosis from apparently normal ECG. We discovered that with ECG alone, the DL model trained from random parameters achieved a specificity of 0.746, the highest among all tested models. This suggests that the DL model trained from scratch can accurately identity no/mild stenosis conditions in the control group using ECG alone. However, its sensitivity was only 0.545, indicating that it might not perform well in predicting severe stenosis in the case group. We hypothesize that adding more features can improve the ability to identify severe cases. By adding patients’ baseline characteristics (age, gender, hypertension, diabetes, dyslipidemia and smoking status), we greatly increased the sensitivity to 0.909. However, the specificity dropped to 0.423, the lowest among the experimental models. This might be due to the lack of training data, causing a high bias in the prediction model.

Therefore, we applied ransfer learning technology and fine-tuned the model based on parameters from an ECG classification task. With ECG-alone transfer learning model, we obtained an acceptable sensitivity of 0.667, but the specificity was still relatively low at 0.577. Finally, we combined ECG and baseline features in the transfer learning model and obtained a relatively high specificity (0.704) as well as a high sensitivity (0.848).

For comparison, we also applied logistic regression using baseline features alone, achieving an accuracy of 0.663 with a sensitivity of 0.545 and sensitivity of 0.718. The baseline characteristics selected (age, gender, hypertension, diabetes, dyslipdemia, smoking status) are generally accepted cardiovascular risk factors [35]. In Table 7, we present the analysis results for these risk factors from the logistic regression. Age, hypertension, and dyslipidemia had significant coefficients, indicating their impact on the discrimination of coronary stenosis severity. This observation is consistent with Table 1, where these factors show small p-values, suggesting a significant association between theses variables and the categories of coronary stenosis (severe or no/mild). However, gender is an exception. While it shows a small p-value in Table 1, it is not considered a significant factor in the logistic regression experiment, which suggests the need for additional investigation. The logistic regression with baseline features had similar performances to the ECG-alone model trained from scratch and outperformed the ECG-alone transfer learning model. This may be due to the fundamental difference between the ECG classification task and the coronary stenosis discrimination task. The model pre-trained for ECC classification may not capture the ECG characteristics that determine the severity of the stenosis. However, by adding the baseline features to the transfer learning model, we can correct the skewed information and achieve the highest performance among the experimental models.

Similar to the results of logistic regression analysis for baseline features, we also report the results of deep features regression analysis in Table 8. Among the deep features, DeepFeature5 (PVC - Premature Ventricular Contraction), DeepFeature9 (VT - Ventricular Tachycardia), DeepFeature11 (AFL - Atrial Flutter), DeepFeature16 (VE - Ventricular Escape), DeepFeature27 (LVH - Left Ventricular Hypertrophy), DeepFeature30 (RAH - Right Atrial Hypertrophy), DeepFeature36 (PACED - Paced Rhythm), and DeepFeature43 (PAUSE - Sinus Pause) have significant coefficients. The findings suggest that these deep features are effective in identifying severe coronary stenosis. However, it is important to clarify that while these deep features correspond to specific ECG classes, they are intermediate representations derived from a pre-trained ECG classification model that has been fine-tuned for the coronary stenosis discrimination task. Consequently, these deep features are no longer associated with diagnosis decisions. Further research is needed to explore the relationship between the pre-trained classification task and the stenosis discrimination task. Currently, the ECG classes should only be regarded as tags for the deep features and should not be overly scrutinized.

Table 8 Regression analysis results for deep and baseline feature
Full size table

In previous AI-based CAD studies using CAG results as labels, the usual criterion for severe stenosis was > 70% lumen diameter stenosis [24, 25]. We selected 90% as the criterion for severe stenosis cases because current guidelines specify that stenoses less than 90% are not necessarily functionally significant (i.e., they do not always induce myocardial ischaemia) [36]. The guidelines also suggest revascularization for lesions with a fractional flow reserve (FFR) ≤ 0.8, which corresponds to coronary stenosis ≥ 90% [37]. Therefore, predicting stenosis ≥ 90% can be used for identifying definite myocardial ischaemia and deciding whether to perform revascularization. Although some patients may have apparently normal ECGs with ≥ 90% coronary stenosis, it is likely that specific ischemia-related changes can be captured by DL models. Compared with other screening tests in medicine, such as the CHA2DS2-VASc score to assess stroke risk in patients with atrial fibrillation (AUC: 0.57 to 0.72) [38], and brain natriuretic peptide (BNP) and N-terminal fragment of proBNP (NT-proBNP) for left ventricular dysfunction detection (AUC: 0.60 and 0.70, respectively) [39], our transfer learning DL model with ECG and baseline features achieved a higher performance, with an AUC of 0.85 for distinguishing ≥ 90% coronary artery stenosis.

This study has some limitations. First, as discussed earlier, the amount of data collected was relatively small, which would have affected the model performance. Future studies should utilize larger datasets to externally validate model performance and examine the effects of baseline features and transfer learning. In addition, we see potential in employing augmentation techniques to simulate a wider variety of conditions and reduce the impact of data scarcity. Second, the CAG results involved in this study lacked sufficient details about coronary lesions, limiting our ability to determine the exact severity of the stenosis, the location of the lesions, and other relevant information. We also grouped patients with no coronary lesion and those with mild coronary lesions into a single category, which restricted the algorithm’s ability to distinguish between various degrees of stenosis. To enhance the algorithm’s capabilities, it is necessary to obtain additional data with more granular classifications. Third, although visualization tools such as gradient-weighted class activation mapping (Grad-CAM) [40] provide intuitive ways to display the feature importance for CNNs, AI ECG models are still considered “black-boxes” lacking interpretability, especially in terms of professional medical explanations. In Figure 7, we show the Grad-CAM mapped ECG examples related to the second-to-last layer of the transfer learning model: A is a TN example; B is an FP example; C is an FN example; and D is a TP example. For the control groups (A and B), the model focused more on the T-P segment, while for the case groups (C and D), redder areas appeared around the QRS complex. However, more nuanced approaches are needed for medical interpretations.

Fig. 7
figure 7

Examples of grad-cAM results

Full size image

For further research, exploring the compatibility between ECG data representations and model structures is also an important direction. When considering ECG data as a one-dimensional signal, 1D CNN models are a suitable choice. Simultaneously, multi-channel ECG data can be represented as a 2D-array, resembling a linear-patterned topographic map. This two-dimensional perspective allows the utilization of 2D models [41]. Furthermore, by applying specific transformation processes [42], we can represent the heart's electrical activity in a three-dimensional space, making it possible to explore 3D models [43]. In future studies, we can compare the results obtained from different data representations and their corresponding models to deepen our understanding and improve predictive accuracy.

Conclusions

Using apparently normal ECGs, we investigated the possibility of employing DL technologies to distinguish severe coronary stenosis (≥ 90%) from no or mild coronary stenosis (< 50%). The manual readings of the selected ECGs did not provide much evidence of the coronary stenosis status. Using a deep CNN trained completely with our labeled data, we achieved an accuracy of 0.683, which is higher than that achieved by cardiologists, but the sensitivity was low. By combining the pre-trained ECG classification model parameters with patients’ baseline characteristics, we achieved an accuracy of 0.750 with a sensitivity of 0.848 and specificity of 0.704. The findings are significant since they not only demonstrate the effectiveness of using AI ECG models to identify severe coronary stenosis, but also validate an effective approach utilizing existing ECG models. With these existing ECG models, we can easily extract “deep features” that summarize the inherent information in the ECG and combine these features with traditional clinical data to obtain valuable auxiliary diagnostic conclusions with minimal computational cost. Our goal is to further optimize the model to achieve early accurate identification of high-risk patients with severe stenosis, facilitating early intervention to reduce mortality and morbidity. At the same time, we aim to improve the interpretability of the model so that medical professionals have more trust in it and promote its application.

Abbreviations

AE:

Atrial escape

AF:

Atrial fibrillation

AFL:

Atrial flutter

AI:

Artificial intelligence

AUC:

Area under the curve

AVBI:

First-Degree atrioventricular block

AVBII:

Second-Degree atrioventricular block

AVBIII:

Third-Degree atrioventricular block

BLV:

Low voltage in limb leads

BNP:

Brain natriuretic peptide

CABG:

Coronary artery bypass grafting

CAD:

Coronary artery disease

CAG:

Coronary angiography

CCW:

Counterclockwise rotation

CLV:

Low Voltage in precordial leads

CNN:

Convolutional neuron network

CW:

Clockwise rotation

DL:

Deep learning

ECG:

Electrocardiograms

ERS:

Early repolarization syndrome

FFR:

Fractional flow reserve

FN:

False negative

FP:

False positive

FPR:

False positive rate

Grad-CAM:

Gradient-weighted class activation mapping

IVB:

Intraventricular block

JE:

Junctional escape

LAD:

Left axis deviation

LAFB:

Left anterior fascicular block

LAH:

Left atrial hypertrophy

LBBB:

Left bundle branch block

LVH:

Left ventricular hypertrophy

LVH:

Left ventricular hypertrophy

MI:

Myocardial infarction

N:

Normal ECG

NT-proBNP:

N-terminal fragment of proBNP

PAC:

Premature atrial contraction

PACED:

Paced rhythm

PAUSE:

Sinus pause

PCI:

Percutaneous coronary intervention

PJC:

Premature junctional contraction

PR:

PR Interval abnormality

PSC:

Premature supraventricular contraction

PVC:

Premature ventricular contraction

Q:

Q Wave abnormality

QCA:

Quantitative coronary angiography

QT:

QT Interval abnormality

RAD:

Right axis deviation

RAH:

Right atrial hypertrophy

RBBB:

Right bundle branch block

ROC:

Receiver operating characteristics

RVH:

Right ventricular hypertrophy

SE:

Supraventricular escape

SN:

Sinus rhythm

SNA:

Sinus arrhythmia

SNB:

Sinus bradycardia

SNT:

Sinus tachycardia

ST:

ST segment abnormality

STTQ:

ST-T-Q abnormality

SVT:

Supraventricular tachycardia

T:

T Wave abnormality

TN:

True negative

TP:

True positive

TPR:

True positive rate

VE:

Ventricular escape

VF:

Ventricular fibrillation

VFL:

Ventricular flutter

VT:

Ventricular tachycardia

WPW:

Wolff-Parkinson-White syndrome

References

  1. Malakar AK, Choudhury D, Halder B, Paul P, Uddin A, Chakraborty S. A review on coronary artery disease, its risk factors, and therapeutics. J Cell Physiol. 2019;234(10):16812–23.

    Article  CAS  PubMed  Google Scholar 

  2. Zhou M, Wang H, Zeng X, Yin P, Zhu J, Chen W, et al. Mortality, morbidity, and risk factors in China and its provinces, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. The Lancet. 2019;394(10204):1145–58.

    Article  Google Scholar 

  3. Van Mieghem CA. CT as gatekeeper of invasive coronary angiography in patients with suspected CAD. Cardiovascular diagnosis and therapy. 2017Apr;7(2):189.

    Article  PubMed  PubMed Central  Google Scholar 

  4. Mahmoodzadeh S, Moazenzadeh M, Rashidinejad H, Sheikhvatan M. Diagnostic performance of electrocardiography in the assessment of significant coronary artery disease and its anatomical size in comparison with coronary angiography. J Res Med Sci. 2011;16(6):750.

    PubMed  PubMed Central  Google Scholar 

  5. Shao C, Wang J, Tian J, Tang YD. Coronary artery disease: From mechanism to clinical practice. Adv Exp Med Biol. 2020;1177:1–36. https://doiorg.publicaciones.saludcastillayleon.es/10.1007/978-981-15-2517-9_1.

  6. Chen X, editor. Huang Wan's clinical electrocardiography (6th ed) (Chinese ed). Beijing: People's Medical Publishing House; 2009. 

  7. Yan Y, Zhang JW, Zang GY, Pu J. The primary use of artificial intelligence in cardiovascular diseases: what kind of potential role does artificial intelligence play in future medicine? Journal of geriatric cardiology: JGC. 2019;16(8):585.

    PubMed  PubMed Central  Google Scholar 

  8. Antoniades C, Oikonomou EK. Artificial intelligence in cardiovascular imaging—principles, expectations, and limitations. Eur Heart J. 2024Apr 14;45(15):1322–6.

    Article  Google Scholar 

  9. Nazir MB, Hussain I. Revolutionizing Cardiac Care: AI and Deep Learning in Heart Health. International Journal of Advanced Engineering Technologies and Innovations. 2024May 7;1(4):189–208.

    Google Scholar 

  10. Clifford GD, Liu C, Moody B, Li-wei HL, Silva I, Li Q, et al. AF classification from a short single lead ECG recording: The PhysioNet/computing in cardiology challenge 2017. In: 2017 Computing in Cardiology (CinC). IEEE; 2017. p. 1-4.

  11. Hong S, Wu M, Zhou Y, Wang Q, Shang J, Li H, ENCASE: An ENsemble ClASsifiEr for ECG classification using expert features and deep neural networks. In, et al. Computing in cardiology (cinc). IEEE. 2017;2017:1–4.

    Google Scholar 

  12. Issa MF, Yousry A, Tuboly G, Juhasz Z, AbuEl-Atta AH, Selim MM. Heartbeat classification based on single lead-II ECG using deep learning. Heliyon. 2023;9(7). https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.heliyon.2023.e17974.

  13. Attia ZI, Kapa S, Lopez-Jimenez F, McKie PM, Ladewig DJ, Satam G, et al. Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram. Nat Med. 2019;25(1):70–4.

    Article  CAS  PubMed  Google Scholar 

  14. Sengupta PP, Kulkarni H, Narula J. Prediction of abnormal myocardial relaxation from signal processed surface ECG. J Am Coll Cardiol. 2018;71(15):1650–60.

    Article  PubMed  Google Scholar 

  15. Lee E, Ito S, Miranda WR, Lopez-Jimenez F, Kane GC, Asirvatham SJ, Noseworthy PA, Friedman PA, Carter RE, Borlaug BA, Attia ZI, Oh JK. Artificial intelligence-enabled ECG for left ventricular diastolic function and filling pressure. npj Digit Med. 2024;7(1):4. https://doiorg.publicaciones.saludcastillayleon.es/10.1038/s41746-023-00993-7.

  16. Cohen-Shelly M, Attia ZI, Friedman PA, Ito S, Essayagh BA, Ko WY, et al. Electrocardiogram screening for aortic valve stenosis using artificial intelligence. Eur Heart J. 2021;42(30):2885–96.

    Article  PubMed  Google Scholar 

  17. Vaid A, Argulian E, Lerakis S, Beaulieu-Jones BK, Krittanawong C, Klang E, Lampert J, Reddy VY, Narula J, Nadkarni GN, Glicksberg BS. Multi-center retrospective cohort study applying deep learning to electrocardiograms to identify left heart valvular dysfunction. Communications Medicine. 2023Feb 14;3(1):24.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Goto S, Mahara K, Beussink-Nelson L, Ikura H, Katsumata Y, Endo J, et al. Artificial intelligence-enabled fully automated detection of cardiac amyloidosis using electrocardiograms and echocardiograms. Nat Commun. 2021;12(1):1–12.

    Article  Google Scholar 

  19. Siontis KC, Wieczorek MA, Maanja M, Hodge DO, Kim HK, Lee HJ, Lee H, Lim J, Park CS, Ariga R, Raman B. Hypertrophic cardiomyopathy detection with artificial intelligence electrocardiography in international cohorts: an external validation study. European Heart Journal-Digital Health. 2024 Apr 15:ztae029.

  20. Raghunath S, Ulloa Cerna AE, Jing L, VanMaanen DP, Stough J, Hartzel DN, et al. Prediction of mortality from 12-lead electrocardiogram voltage data using a deep neural network. Nat Med. 2020;26(6):886–91.

    Article  CAS  PubMed  Google Scholar 

  21. Hughes JW, Tooley J, Torres Soto J, Ostropolets A, Poterucha T, Christensen MK, Yuan N, Ehlert B, Kaur D, Kang G, Rogers A. A deep learning-based electrocardiogram risk score for long term cardiovascular death and disease. NPJ digital medicine. 2023Sep 12;6(1):169.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Hong S, Zhou Y, Shang J, Xiao C, Sun J. Opportunities and challenges of deep learning methods for electrocardiogram data: A systematic review. Comput Biol Med. 2020;122: 103801.

    Article  PubMed  Google Scholar 

  23. Holmstrom L, Christensen M, Yuan N, Weston Hughes J, Theurer J, Jujjavarapu M, Fatehi P, Kwan A, Sandhu RK, Ebinger J, Cheng S. Deep learning-based electrocardiographic screening for chronic kidney disease. Communications Medicine. 2023May 26;3(1):73.

    Article  PubMed  PubMed Central  Google Scholar 

  24. Leasure M, Jain U, Butchy A, Otten J, Covalesky VA, McCormick D, et al. Deep learning algorithm predicts angiographic coronary artery disease in stable patients using only a standard 12-lead electrocardiogram. Can J Cardiol. 2021;37(11):1715–24.

    Article  PubMed  Google Scholar 

  25. Huang PS, Tseng YH, Tsai CF, Chen JJ, Yang SC, Chiu FC, et al. An Artificial Intelligence-Enabled ECG Algorithm for the Prediction and Localization of Angiography-Proven Coronary Artery Disease. Biomedicines. 2022;10(2):394.

    Article  PubMed  PubMed Central  Google Scholar 

  26. Tang P, Wang Q, Ouyang H, Yang S, Hua P. The feasibility of early detecting coronary artery disease using deep learning-based algorithm based on electrocardiography. Aging (Albany NY). 2023May 5;15(9):3524.

    Article  PubMed  Google Scholar 

  27. Lee YH, Hsieh MT, Chang CC, Tsai YL, Chou RH, Lu HHS, et al. Improving detection of obstructive coronary artery disease with an artificial intelligence-enabled electrocardiogram algorithm. Atherosclerosis. 2023;381: 117238.

    Article  CAS  PubMed  Google Scholar 

  28. Belmonte M, Paolisso P, Gallinoro E, Bertolone DT, Leone A, Esposito G, Caglioni S, Viscusi MM, Bermpeis K, Storozhenko T, Wyffels E, Bartunek J, Sonck J, Collet C, Andreini D, Vanderheyden M, Penicka M, Barbato E. Improved diagnostic accuracy of vessel-specific myocardial ischemia by coronary computed tomography angiography (CCTA). J Cardiovasc Comput Tomogr. 2024;S1934–5925(24):00449–0. https://doiorg.publicaciones.saludcastillayleon.es/10.1016/j.jcct.2024.09.015. Epub ahead of print.

  29. Liu X, Li C, Zhang Q, Meng Q, Zhang H, Li Z. Comparative study of myocardial perfusion and coronary flow velocity reserve derived from adenosine triphosphate stress myocardial contrast echocardiography in coronary lesions with no/mild stenosis. Frontiers in Cardiovascular Medicine. 2024;11:1353736.

    Article  PubMed  PubMed Central  Google Scholar 

  30. Harmon KG, Zigman M, Drezner JA. The effectiveness of screening history, physical exam, and ECG to detect potentially lethal cardiac disorders in athletes: a systematic review/meta-analysis. J Electrocardiol. 2015May 1;48(3):329–38.

    Article  PubMed  Google Scholar 

  31. Hong S, Fu Z, Zhou R, Yu J, Li Y, Wang K, et al. Cardiolearn: a cloud deep learning service for cardiac disease detection from electrocardiogram. In: Companion Proceedings of the Web Conference 2020; 2020. p. 148–52.

  32. Hong S, Xu Y, Khare A, Priambada S, Maher K, Aljiffry A, et al. HOLMES: Health OnLine Model Ensemble Serving for Deep Learning Models in Intensive Care Units. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining; 2020. p. 1614–24.

  33. Fu Z, Hong S, Zhang R, Du S. Artificial-intelligence-enhanced mobile system for cardiovascular health management. Sensors. 2021;21(3):773.

    Article  PubMed  PubMed Central  Google Scholar 

  34. Boateng EY, Abaye DA. A review of the logistic regression model with emphasis on medical research. Journal of data analysis and information processing. 2019Oct 12;7(04):190.

    Article  Google Scholar 

  35. Kannel WB, McGee DL. Diabetes and cardiovascular risk factors: the Framingham study. Circulation. 1979;59(1):8–13.

    Article  CAS  PubMed  Google Scholar 

  36. Knuuti J, Wijns W, Saraste A, Capodanno D, Barbato E, Funck-Brentano C, et al. 2019 ESC Guidelines for the diagnosis and management of chronic coronary syndromes: the Task Force for the diagnosis and management of chronic coronary syndromes of the European Society of Cardiology (ESC). Eur Heart J. 2020;41(3):40777.

    Article  Google Scholar 

  37. Toth G, Hamilos M, Pyxaras S, Mangiacapra F, Nelis O, De Vroey F, et al. Evolving concepts of angiogram: fractional flow reserve discordances in 4000 coronary stenoses. Eur Heart J. 2014;35(40):2831–8.

    Article  PubMed  Google Scholar 

  38. Wu JT, Wang SL, Chu YJ, Long DY, Dong JZ, Fan XW, Yang HT, Duan HY, Yan LJ, Qian P. CHADS2 and CHA2DS2-VASc scores predict the risk of ischemic stroke outcome in patients with interatrial block without atrial fibrillation. J Atheroscler Thromb. 2017;24(2):176–84. https://doiorg.publicaciones.saludcastillayleon.es/10.5551/jat.34900.

  39. Bhalla V, Isakson S, Bhalla MA, Lin JP, Clopton P, Gardetto N, et al. Diagnostic ability of B-type natriuretic peptide and impedance cardiography: testing to identify left ventricular dysfunction in hypertensive patients. Am J Hypertens. 2005;18(S2):73S–81S.

    Article  CAS  PubMed  Google Scholar 

  40. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-cam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision; 2017. p. 618–26.

  41. Zhang D, Geng S, Zhou Y, Xu W, Wei G, Wang K, et al. Artificial Intelligence System for Detection and Screening of Cardiac Abnormalities using Electrocardiogram Images. arXiv preprint arXiv:2302.10301 . 2023 Feb 10.

  42. Dawson D, Yang H, Malshe M, Bukkapatnam ST, Benjamin B, Komanduri R. Linear affine transformations between 3-lead (Frank XYZ leads) vectorcardiogram and 12-lead electrocardiogram signals. J Electrocardiol. 2009Nov 1;42(6):622–30.

    Article  PubMed  Google Scholar 

  43. Goenka N, Tiwari S. AlzVNet: A volumetric convolutional neural network for multiclass classification of Alzheimer’s disease through multiple neuroimaging computational approaches. Biomed Signal Process Control. 2022Apr;1(74): 103500.

    Article  Google Scholar 

Download references

Acknowledgements

We would like to thank the editors and reviewers for their thorough evaluation and insightful comments. Their feedback has significantly contributed to enhancing the quality and clarity of this manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (No. 62102008), Clinical Medicine Plus X—Young Scholars Project of Peking University, the Fundamental Research Funds for the Central Universities (PKU2024LCXQ030), PKU-OPPO Fund (B0202301), the Tianjin Municipal Natural Science Foundation (No. 21JCZDJC01080), Tianjin Key Medical Discipline (Specialty) Construction Project (No. TJYXZDXK-029A), and the National Natural Science Foundation of China (No. 82470527).

The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Author information

Author notes

  1. Zhengkai Xue and Shijia Geng contributed equally to this work.

Authors and Affiliations

  1. Tianjin Key Laboratory of Ionic-Molecular Function of Cardiovascular Disease, Department of Cardiology, Tianjin Institute of Cardiology, The Second Hospital of Tianjin Medical University, Tianjin, China

    Zhengkai Xue, Sutao Hu & Kangyin Chen

  2. HeartVoice Medical Technology, Hefei, China

    Shijia Geng, Deyun Zhang & Weilun Xu

  3. Department of Cardiology, Cardiovascular Center, Beijing Friendship Hospital, Capital Medical University, Beijing, China

    Shaohua Guo

  4. Department of Echocardiography, The First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, China

    Guanyu Mu

  5. Department of Cardiology, Tianjin Union Medical Center, Nankai University Affiliated Hospital, Tianjin, China

    Bo Yu & Peng Wang

  6. Heart Center, Tianjin Third Central Hospital, Tianjin, China

    Yanhong Liu & Lei Yang

  7. Tianjin Key Laboratory of Extracorporeal Life Support for Critical Diseases, Artificial Cell Engineering Technology Research Center, Tianjin Institute of Hepatobiliary Disease, Tianjin, China

    Yanhong Liu

  8. The Second Hospital of Tianjin Medical University, Tianjin, China

    Huayue Tao

  9. National Institute of Health Data Science, Peking University, Beijing, China

    Shenda Hong

  10. Institute of Medical Technology, Peking University Health Science Center, Beijing, China

    Shenda Hong

  11. Institute for Artificial Intelligence, Peking University, Beijing, China

    Shenda Hong

  12. Department of Emergency Medicine, Peking University First Hospital, Beijing, China

    Shenda Hong

  13. School of Precision Instrument and Opto-Electronic Engineering, Tianjin University, Tianjin, China

    Kangyin Chen

Authors
  1. Zhengkai Xue
    View author publications

    You can also search for this author inPubMed Google Scholar

  2. Shijia Geng
    View author publications

    You can also search for this author inPubMed Google Scholar

  3. Shaohua Guo
    View author publications

    You can also search for this author inPubMed Google Scholar

  4. Guanyu Mu
    View author publications

    You can also search for this author inPubMed Google Scholar

  5. Bo Yu
    View author publications

    You can also search for this author inPubMed Google Scholar

  6. Peng Wang
    View author publications

    You can also search for this author inPubMed Google Scholar

  7. Sutao Hu
    View author publications

    You can also search for this author inPubMed Google Scholar

  8. Deyun Zhang
    View author publications

    You can also search for this author inPubMed Google Scholar

  9. Weilun Xu
    View author publications

    You can also search for this author inPubMed Google Scholar

  10. Yanhong Liu
    View author publications

    You can also search for this author inPubMed Google Scholar

  11. Lei Yang
    View author publications

    You can also search for this author inPubMed Google Scholar

  12. Huayue Tao
    View author publications

    You can also search for this author inPubMed Google Scholar

  13. Shenda Hong
    View author publications

    You can also search for this author inPubMed Google Scholar

  14. Kangyin Chen
    View author publications

    You can also search for this author inPubMed Google Scholar

Contributions

ZX prepared the data, SG did the algorithm development, and ZX and SG wrote the main manuscript text. SGu and GM prepared the figures and did the review. BY, PW and SHu provided the statistical results and prepared the tables. DZ and WX did the data curation and algorithm verification. YL and LY worked on the conceptualization. HT, SH and KC substantively revised the paper review and guided on the design of the work. All authors reviewed the manuscript.

Corresponding authors

Correspondence to Huayue Tao, Shenda Hong or Kangyin Chen.

Ethics declarations

Ethics approval and consent to participate

All methods were carried out in accordance with relevant guidelines and regulations. All experimental protocols were approved by the ethics committee of the Second Hospital of Tianjin Medical University (Tianjin, China). Informed consent was obtained from all subjects.

Consent for publication

Not applicable.

Data availability

Data used for this study are not publicly available due to privacy concerns. Part of the data is available upon request to the corresponding author. The model code is available at https://github.com/hsd1503/resnet1d.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Xue, Z., Geng, S., Guo, S. et al. Screening for severe coronary stenosis in patients with apparently normal electrocardiograms based on deep learning. BMC Med Inform Decis Mak 24, 355 (2024). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12911-024-02764-0

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12911-024-02764-0

Keywords

BMC Medical Informatics and Decision Making

ISSN: 1472-6947

Contact us