Four-Channel ECG as a Single Source for Early Diagnosis of Cardiac Hypertrophy and Dilation — A Deep Learning Approach

The ECG data are in standard 10 s, 12-lead format with a sampling rate of 500 Hz, and from three campuses (Main Campus, Optical Valley Campus, and Sino-French New City Campus) of Tongji Hospital. The majority of data were acquired using a General Electric (GE) Marquette MAC 3500 or 5500 ECG machine (GE and Marquette University, WI, United States), and Philips PageWriter TC30 and TC50 machines (Philips, Shenzhen Goldway Industrial Inc., Shenzhen, China), while a minority were acquired using the mecg1000 (MedeX Technologies Limited, Beijing, China). The ECG data from the First People’s Hospital of Jiangxia District, Wuhan, China (abbreviated as Jiangxia Hospital), were acquired using different machine brands, including a GE Marquette MAC 800 ECG machine (GE and Marquette University, WI, United States), and mecg300 and mecg200 (MedeX Technologies Limited, Beijing, China). Echocardiography data were acquired by a GE Vivid E9 ultrasound scanner (GE Vingmed Ultrasound, Horten, Norway) for all patients. Offline measurements were performed by an experienced cardiologist using EchoPAC (version: 203.66; GE Vingmed Ultrasound, Horten, Norway). Diastolic interventricular septum (IVS) thickness, left ventricle (LV) posterior wall (PW) thickness, LV end-diastolic dimensions, right ventricle (RV) end-diastolic dimensions, and left atrium (LA) end-systolic dimensions were measured from the parasternal long-axis view, and right atrium (RA) systolic dimensions were measured from the apical four-chamber view. According to the guidelines, LVH/LVD was further classified as a diameter of greater than 53 mm in men and greater than 51 mm in women and/or IVS and LV PW diameter of greater than 10 mm. RVH/RVD was further classified as a diameter of greater than 33 mm in men and women, LAE as a diameter of greater than 37 mm in men and greater than 35 mm in women, and RAE as a diameter of greater than 45 mm in men and women.^26,27

In our dataset, each ECG record is associated with a gold standard with respect to its corresponding echocardiography record. For example, a patient with a normal ECG can be labeled as having LVH based on the echocardiography results, whereas a patient diagnosed with a single LVH label by ECG may also have RVH according to additional echocardiography examination results. Only one patient identifier (ID) was assigned to each patient during the visit. This ensured that the patient had the same ID when undergoing the ECG and echocardiography tests.

We use an open-access dataset (PTB-XL) to examine the generalizability of the proposed approach.²⁸ The PTB-XL ECG dataset is a dataset of 21,837 clinical 12-lead ECGs from 18,885 patients of 10 s length. The diagnoses within this dataset are categorized into five superclasses, including normal ECG, myocardial infarction, ST/T change, conduction disturbance, and hypertrophy. Each superclass contains multiple subclasses. Specifically, for hypertrophy, the PTB-XL dataset provides LVH, left atrial overload/enlargement (LAO/LAE), RVH, right atrial overload/enlargement (RAO/RAE), and septal hypertrophy (SEHYP). For our tests, we focus on data labeled under the normal ECG and hypertrophy superclasses and utilize the provided standard communications protocol for computer-assisted electrocardiography (SCP-ECG) statements as the gold standard. The SCP-ECG statements labeled each subclass to some extent and, since there is no further annotation, we selected the records labeled with the code “100.” To be consistent with our multi-label five-class model, ECGs with SEHYP labels are classified as LVH. Our data preprocessing results in a test dataset comprising 8081 12-lead ECGs, including 151 LAOs/LAEs (referred to as the LAE group), 776 LVHs and SEHYPs (referred to as the LVH and/or LVD group), 33 RAOs/RAEs (referred to as the RAE group), 28 RAOs/RAEs (referred to as the RVH and/or RVD group), and 7182 normal classes.

Three ECG physicians working in the cardiology departments of Tongji Hospital and three ECG physicians working in the cardiology departments of Jiangxia Hospital were invited to explain the ECGs to assess the comparative performance of the deep learning model. We divided the physicians into two groups according to their experience of working with ECGs: 0 to 6 years and 6 to 24 years. Table S1 in the Supplementary Appendix shows the number of physicians and the mean age, sex information, and experience level of each group. Each physician received 473 ECGs from Tongji Hospital and was informed that only the five classes included in this study could be selected (either singly or in combination).

To detect whether a patient has one or more of the four types of diseases (LVH and/or LVD, RVH and/or RVD, LAE, and RAE), or is a patient without any of these conditions (normal), we have formulated a multi-label diagnostic model, denoted ECG-Cor-Net (model architecture and parameters are presented in Fig. S2 and Table S6, respectively), to detect multiple types of LVH and/or LVD, RVH and/or RVD, LAE, and RAE with ECG as a single source. Each ECG input X corresponded to an output vector $Y\in {ℝ}^{1\times 5}$. Our aim is to learn a model as a nonlinear function $\tilde{Y}=f(x)$ to minimize the cost function between the ground-truth label Y and the predicted label $\tilde{Y}$. We propose a novel aggregated attribution score for evaluating ECG data pertaining to different classes. This score is computed for each lead and time point based on the expected gradients of the ECG-Cor-Net model, incorporating concepts from integrated gradients, a gradient-based pixel attribution method, and Shapley additive explanations, an additive feature attribution method.²⁹ The binary classification model, denoted as $F:{ℝ}^{5000\times 12}\to [0,1]$, represents a nonlinear function to diagnose one of the abnormalities of LVH and/or LVD, RVH and/or RVD, LAE, and RAE.

The proposed ECG-Cor-Net model is a multi-label model composed of a one-dimensional (1D) flow and a two-dimensional (2D) flow that correspond to the original 10 s time series and the segmented images of the 12-channel input, respectively, as shown in the blue and red halves in Figure S2. The segmented images of the 12-channel model are one heartbeat of a 10 s ECG sample. We cropped out the flattened period (first and last 40 samples) of the segmented waveform.³⁰ The cropped waveform samples were then converted into a 128 × 128 resolution 2D image of each lead as the input of 2D flow (see heartbeat segmentation details in Fig. S2). The overall architecture of the ECG-Cor-Net model is stacked by four modules, namely a 2D image conversion module (in the 2D flow only) to investigate the rich features of morphological characteristics from image segmentations of the ECG input, an embedding module to extract hidden features in the latent space, a correlation module to explore interdependencies between different channels and spatial locations of the extracted hidden features, and a multi-label diagnosis module for the diagnosis of five classes. See the parameters of each module in Table S6. We choose four leads with high aggregated attribution scores to obtain two simplified models for the detection of cardiac hypertrophy, dilation, and enlargement. In our experiments, we compare the performance of the simplified four-channel models and the 12-channel model. We denote the basic model, trained only using 1D time series data without utilizing 2D images, as ECG-Cor-Net-1D. We also compare the performance of ECG-Cor-Net and ECG-Cor-Net-1D to demonstrate the importance of the implementation of 2D images as another representation.

We use multi- and single-label validation metrics to verify the performance of the proposed ECG-Cor-Net model. For multi-label validation, five diagnostic metrics are utilized, including subset accuracy,³¹ Hamming loss,³² average F1-score, average recall, and average precision. Subset accuracy measures the exact match ratio for the multi-label diagnosis problem, indicating the percentage of accurately predicted ECGs for all five labels. Hamming loss, another multi-label metric, assesses the Hamming distance between predicted and true labels, penalizing individual label discrepancies. Single-label metrics, such as area under (AU) receiver operating characteristic curve (AUROC), AU precision-recall curve (AUPRC), sensitivity, specificity, F1-score, and accuracy, are calculated for different classes. The calculation formulas for each metric can be found in Section S4.

Our cardiac hypertrophy, dilation, and enlargement dataset was based on the complete standard ECG database of the three campuses of Tongji Hospital (Main Campus, Optical Valley Campus, and Sino-French New City Campus) at Huazhong University of Science and Technology, People’s Republic of China. Ethical approval was obtained from Tongji Hospital (No. TJ-IRB20220553). As shown in Figure 2, from January 1, 2012 to July 17, 2021, there were a total of 2,386,886 ECGs and 988,257 echocardiograms recorded. We identified echocardiograms categorized under cardiac hypertrophy, dilation, enlargement, and normal, and matched them to the corresponding ECGs. To construct a balanced dataset, we identified 90,895 ECGs from 74,562 inpatients and outpatients, who underwent both ECG and echocardiography examinations at one visit or within a short time frame. A total of 7377 patients were excluded if they were less than18 years of age. A total of 541 ECGs were excluded because of detached electrodes. After exclusions, 80,007 ECGs from 65,927 patients (57.25±14.38 years, 59% men) remained for cross-validation training (the exclusion flowchart is shown in Fig. 2). As for the training dataset, we have included all the echocardiography-matched ECGs without considering ECG abnormalities. However, for the test datasets we have excluded abnormalities related to depolarization from our analysis (e.g., bundle branch blocks, ventricular pacing), as these data would be reviewed by physicians who could be influenced by such depolarization abnormalities. We divided the cross-validation dataset for training, validation, and hold-out testing (denoted as the Hold-out Test), based on patient IDs, at a ratio of 8:1:1. The label distributions of each class in the cross-validation training dataset are presented Figure S1A. An external test dataset from Tongji Hospital (denoted as the TJ-Test), containing 473 ECGs from 469 new patients was constructed for external validation (see the label distributions of each class in Fig. S1B). To assess the generalization ability of the ECG-Cor-Net model, we built another external dataset from Jiangxia Hospital (denoted as the JX-Test), containing 177 ECGs from 166 patients in Wuhan, China, to verify the performance of the model in multicenter hospitals. The age and sex information of each class in the cross-validation training, TJ-Test set, and JX-Test set are presented in Table S5. We also use the public PTB-XL dataset to examine the generalizability of the proposed approach, which contains 8081 clinical 12-lead ECGs.

We have analyzed the cross-validation training dataset and the TJ-Test set, and collected the patients’ electronic health records (EHRs), as shown in Table S13. There is a higher number of inpatients than outpatients in the cross-validation dataset, whereas the TJ-Test set displays a similar ratio of inpatients to outpatients. Most patients are coming from various provinces of China, including Hubei, Henan, and Anhui. In terms of etiologies, congenital cardiovascular diseases rank first in the RVH and/or RVD (28.1%) and RAE (15.1%) groups, hypertension ranks first in the LVH and/or LVD (5.5%) group, and valvular diseases rank first (7.9%), and hypertension ranks second (5.4%) in the LAE group. Regarding comorbidities, coronary heart disease is the most prevalent in patients with LVH and/or LVD (6.8%), RVH and/or RVD (5.7%), and LAE (13.9%). Atrial fibrillation ranks first in patients with RAE (8.4%). Furthermore, there are 24,933/80,007 (31.2%) patients with LVH and/or LVD in the cross-validation training dataset. This ratio is consistent with the ranges reported in these studies.³³

All ECGs are classified into five specific classes according to their echocardiographic findings: normal, LVH and/or LVD, RVH and/or RVD, LAE, and RAE. After exclusion, there were 5493, 11,023, 24,933, and 37,153 ECG samples for RVH and/or RVD, RAE, LVH and/or LVD, and LAE, respectively, remaining for analysis (Fig. 3B). The identified patients (except normal) had at least one abnormality in the left and/or right atriums and/or ventricles. A multi-label ECG summary of each class in the cross-validation training dataset is presented in Figure 3 (Fig. S1A shows the overall label distributions). Overall, ECGs diagnosed with LAE account for 46% (37,153/80,007) of the overall abnormal samples, and LVH and/or LVD account for almost 31% (24,933/80,007), indicating that most cardiac hypertrophy, dilation, and enlargement occurs in the left atrium and ventricle. Figure 3B shows that RVH and/or RVD and RAE have high multi-label percentages of 98% and 75%, respectively, indicating that patients with RVH and/or RVD or RAE often have functional abnormalities in other areas of the heart. LVH and/or LVD and LAE had relatively low multi-label percentages (36% and 31%, respectively). In addition, Figure S1A shows that a large proportion of RVH abnormalities (94.8%; 5209/5493) are also diagnosed as RAE, suggesting that enlargement of the RV is highly correlated with an increase in the size of the RA. Multiple labeling at such a high rate confirms the significantly high prevalence of LVH and/or LVD, RVH and/or RVD, LAE, and RAE, and complicates the training of the model. The average values and standard deviations of the key indicators for diagnosing cardiac hypertrophy, dilation, and enlargement, such as LV, RV, LA, RA, and aortic root diameters, are presented in Tables S2 and S4 for the multi-label and single-label patients, respectively. The missing values within these indicators are also statistically accounted for in Table S3.

We conducted nine rounds of cross-validation training by randomly splitting the first 90% (i.e., the first 8:1 among 8:1:1) for training and validating. This training process is repeated eight times, so nine different models are obtained for performance assessment on all test sets. The detection results of the ECG-Cor-Net model on the external TJ-Test set are compared with those of six experienced ECG physicians, categorized into two groups according to their working experience (9.67±6.26 years), i.e., physicians one to three have 0 to 6 years of experience and physicians four to six have 6 to 24 years of experience. The AUROC and AUPRC curves corresponding to the single-label metrics of RVH and/or RVD, RAE, and normal are shown in Figure 4 and Table 1, whereas the results of all five classes are shown in Figure S3 and Figure S4. The details of the participating physicians are provided in Table S1. Consistent with clinical practice, in Figure 4A, physician markers are grouped at the lower left corner of the plots for RVH and/or RVD and RAE, indicating the low sensitivity (associated with the probability of a correct diagnosis conditioned on truly being hypertrophy/dilation of the heart) and high specificity (associated with the probability of a correct diagnosis conditioned on truly being nonhypertrophy/nondilation of the heart) of clinical ECG readings on cardiac hypertrophy, dilation, and enlargement. In contrast, the ECG-Cor-Net model achieved high AUROC values with LAE, LVH and/or LVD, RAE, RVH and/or RVD, and normal values of 0.837, 0.881, 0.888, 0.898, and 0.929, respectively. Similarly, in Figure 4B, physician markers are grouped to the left of the abnormalities, indicating low recall values (similar to the sensitivity values in Fig. 4A) of clinical ECG readings on cardiac hypertrophy, dilation, and enlargement.

We found that the physician groups with 0 to 6 and 6 to 24 years of experience both obtained sensitivities of RAE and RVH and/or RVD even lower than 0.100 (0.018±0.031 and 0.029±0.013 for 0 to 6 year group, and 0.036±0.047 and 0.065±0.038 for 6 to 24 year group, respectively), confirming that the rate of missed diagnoses on ECG interpretation in clinical practice is extremely high. Fortunately, the ECG-Cor-Net model significantly increases diagnostic performance compared with physician performance. The model diagnostic scores of the AUROCs (Fig. 4A) and AUPRCs (Fig. 4B) are significantly higher than those of the physicians. F1-score is increased from physicians’ 0.159±0.173 for LAE, 0.321±0.128 for LVH and/or LVD, 0.044±0.059 for RAE, 0.080±0.046 for RVH and/or RVD, and 0.692±0.012 for normal, to the model’s 0.655±0.039, 0.614±0.044, 0.521±0.039, 0.531±0.038, and 0.855±0.005, respectively; sensitivity of abnormal classes is increased from physicians’ 0.111±0.149 for LAE, 0.224±0.131 for LVH and/or LVD, 0.027±0.037 for RAE, and 0.047±0.032 for RVH and/or RVD, to the model’s 0.612±0.064, 0.565±0.080, 0.417±0.042, and 0.440±0.052, respectively. The results of our ECG-Cor-Net model are better than those of the ECG-Cor-Net-1D model (without 2D image input) (Fig. 4, Figs. S3 and S4).

More importantly, we examined the multi-label diagnostic results using five multi-label diagnostic metrics: subset accuracy, Hamming loss, average F1-score, average recall, and average precision. A comparison of the proposed ECG-Cor-Net with six experienced physicians in terms of their multi-label diagnostic performance is presented in Figure 5. Typically, subset accuracy is a stringent metric that requires a precise match between the predicted and true labels, and which equally penalizes all predictions regardless of whether they are nearly accurate or entirely incorrect. Hamming loss is employed to count instances of label misclassification. The ECG-Cor-Net model results on the TJ-Test set outperform those of all the physicians. We further compared the subset accuracy results of the ECG-Cor-Net and ECG-Cor-Net-1D models with those of physicians (see Fig. S5A). The results improved from physicians’ 0.491±0.017 to ECG-Cor-Net-1D’s 0.569±0.016, further increasing to ECG-Cor-Net’s 0.592±0.016. We further validated the model’s performance on data from different ECG devices in the test set and provided experimental results for the two most common devices (anonymized as manufacturer 1 and manufacturer 2). The model performance shows similar results for ECGs from different manufacturers in diagnosing the above abnormalities (see results in Table S7).

To assess the performance of the proposed model in multicenter hospitals, the proposed ECG-Cor-Net model was externally tested on the JX-Test dataset, and the results are presented in Table S8. We place significant emphasis on the sensitivity and F1-score results, similarly to tests on the TJ-Test set, as these values tend to be consistently low in real-world clinical diagnoses owing to the limited cardiac hypertrophy, dilation, and enlargement characteristics of ECGs, leading to frequent misdiagnoses by ECG diagnostic physicians. Except for the limited availability of RVH and/or RVD samples (the record of only one patient was available) from the external JX-Test set, other experimental findings validate the efficacy of our proposed ECG-Cor-Net model in diagnosing LVH and/or LVD, RVH and/or RVD, LAE, and RAE on ECGs obtained from another hospital.

During cross-validation training, as shown in Figure 2, the last 10% of the patient IDs were fixed and not used for training and validation, serving as the Hold-out Test set. The average performances of the different validation metrics are presented in Table S9. These results closely align with the test set results of the TJ-Test set in Table 1 and JX-Test set in Table S8, further confirming the appropriateness of our data partitioning approach and the model’s robust generalization capabilities.

Next, we investigate the highest-performing leads for detecting cardiac hypertrophy, dilation, and enlargement. This analysis aims to provide ECG physicians with more definitive guidance, building upon the results obtained from the deep learning–based ECG-Cor-Net model with a black-box nature. To this end, we propose an aggregated attribution score for each lead, l, of each class, based on the expected gradient of the proposed ECG-Cor-Net at each time point. The higher the score, the more significant the lead, l, was for the correct prediction of that class. To avoid the inference of ECG properties between different types of abnormalities, we use the four trained binary classification models for the calculation of aggregated attribution scores.

The values of each lead correspond to each class, and the average values of the four classes are shown in Figure 6. It has been suggested that leads I, II, and aVR in the limb leads and V1, V5, and V6 in the precordial leads are more important in the diagnosis of RVH and/or RVD than other leads. In the case of RAE, the scores for leads I and V1 reach 0.161 and 0.108, respectively, ranking as the top two leads in diagnosing RAE. For LVH and/or LVD and LAE, the scores for leads I, II, and aVR in the limb leads and V1, V2, and V4 in the precordial leads are all notably high. In Figure 6, we can see that the top three representative leads with the highest scores are leads I/II in the limb leads and V1, followed by V5/V6 in the precordial leads.

To further validate the effectiveness of the five representative leads identified, we retrained the model with only four input leads chosen from the five representative leads. Using the same training dataset, we obtained two simplified models: one with the inputs of leads I, aVR, V1, and V5, named ECG-Cor-Net-rep-I, and another with the inputs of leads II, aVR, V1, and V5, named ECG-Cor-Net-rep-II. We tested the simplified models on the TJ-Test set, and the subset accuracy results, compared with the original ECG-Cor-Net model, are shown in Figure 7 (subset accuracy results and their 95% confidence intervals [CIs]). The AUROC and AUPRC values compared with the ECG-Cor-Net and ECG-Cor-Net-1D models are shown in Figure S6 and Figure S7, respectively. In terms of subset accuracy, we observe that the subset accuracy of the two simplified models, ECG-Cor-Net-rep-I and ECG-Cor-Net-rep-II, is slightly lower than that of the original ECG-Cor-Net model, decreasing from 0.592 to 0.551 and 0.534, respectively. When compared with physicians, both models maintained a very high level of accuracy. Specifically, ECG-Cor-Net-rep-II shows an average sensitivity of 0.554, specificity of 0.894, F1-score of 0.603, and accuracy of 0.833 in diagnosing heart structure abnormalities and the normal group; detailed information is shown in Table S12. The AUROC and AUPRC scores of the two simplified models are comparable to those of the ECG-Cor-Net and ECG-Cor-Net-1D models.