Machine Learning Estimation of Myocardial Ischemia Severity Using Body Surface ECG

Abstract

Acute myocardial ischemia (AMI) is one of the leading causes of cardiovascular deaths around the globe. Yet, clinical early detection and patient risk stratification of AMI remain an unmet need, in part due to poor performance of traditional electrocardiogram (ECG) interpretation. Machine learning (ML) techniques have shown promise in analysis of ECGs, even detecting cardiac diseases not identifiable via traditional analysis. However, there has been limited usage of ML tools in the case of AMI due to a lack of high-quality training data, especially detailed ECG recordings throughout the evolution of ischemic events. In this study, we applied ML to predict the ischemic tissue volume directly from body surface ECGs in an AMI animal model. The developed ML networks performed favorably, with an average R² value of 0.932 suggesting a robust prediction. The study also provides insights on how to create and utilize ML tools to enhance clinical risk stratification of patients experiencing AMI.

1. Introduction

Acute myocardial ischemia (AMI) is one of the leading causes of sudden cardiac death worldwide.[1, 2] AMI results from an abrupt imbalance between cardiac metabolic demand and cardiac perfusion, leading to severe malfunctions of the heart.[3–5] There remains a critical yet unmet need for consistent, early, clinical detection and risk stratification of patients who might experience severe morbidity or even mortality due to AMI.[5] The task is fairly challenging, especially when patients show atypical symptoms of complete coronary obstruction, thereby calling for a more useful and clinically adaptable tool to facilitate better diagnoses and prognoses.

The electrocardiogram (ECG) is a noninvasive tool commonly used to diagnose and monitor AMI.[6] Even so, the use of ECG-based tools in risk stratification is limited, since clinical ECG diagnosis does not provide precise information about the size of ischemic regions in the heart.[7,8] On the other hand, machine learning (ML) techniques have been proven to be a useful and effective supportive tool to identify cardiac disorders otherwise undetectable in traditional ECG interpretation techniques.[9]

Contemporary application of ML in predicting cardiovascular diseases is limited, primarily due to the lack of high-resolution, controlled, and well-labeled training datasets.[10] Specifically, the limited sample size can introduce biases in the trained networks, making them more difficult to implement in broader clinical scenarios.[10] Large animal model data, however, may provide a viable avenue to develop such ML analytic tools, due to the detail, volume, and control in the data available from such models.[8]

In this study, we leverage a unique dataset of large animal experimental recordings of myocardial ischemic events to create robust ML tools for myocardial ischemia evaluation. We develop ML networks that can predict the size of the ischemic regions directly from the body surface ECG recordings. This study aims to demonstrate a step toward the improvement of ML-ECG tools for clinical risk stratification of AMI patients by leveraging animal data.

2. Methods

Dataset:

Our dataset consisted of continuous recordings of controlled graded myocardial ischemia in a porcine large animal model described in Zenger et al.[8, 11, 12] The dataset consisted of 8 replicates in which recordings were taken on the body surface and heart surface, as well as within intracardiac space during controlled episodes of ischemic stress. During ischemic interventions, the blood flow through the left anterior descending (LAD) coronary artery was controlled by a surgically implanted hydraulic occluder. Reversible transient ischemic stress was induced by a combination of LAD occlusion (up to 90%), and either atrial pacing or pharmacological stimulation via dobutamine. The recordings were made continuously and simultaneously across all electrodes at 1,000 Hz. Each experiment consisted of 3 to 9 individual ischemic interventions, as well as the corresponding baseline and recovery control recordings before and after each intervention. The signal recording and processing were completed by a suite of custom processing tools, hardware, and software.[8,13]

The individual heartbeats from each experiment were isolated using PFEIFER, and resampled to 1,000 time steps[13]. Ischemic tissue volume was calculated using the transmural plunge needle recordings as described previously.[11] Briefly, needle geometry was acquired using postexperiment magnetic resonance imaging (MRI) processes. Then, for each heartbeat, the ST segment potentials were interpolated into this volume. The volume of tissues that had an ST segment deviation above 3 mV was considered ischemic, especially the value of ST40%, the measured voltage at 40% of the ST segment.[8] The ischemic volume is the gold standard that the ML-ECG models target to achieve (Figure 1).

Figure 1.

Sample Visualization of the input and output of the ML tools. The ML network predicts the ischemic tissue volumes (top) from the torso surface ECG recordings (bottom, ST40% values interpolated). The intracardiac needle electrodes recording the electrograms were visualized within the heart geometry (gray spheres). Based on the error of the specific prediction, the minimum (left) and maximum (right) error cases were found and the input-output pairs were visualized, respectively. The ST40% values of the best and worst performing heartbeat were used to visualize across the torso geometry. For demonstration purposes, the best-performing network #3 was used for the error evaluation here.

In total, 79,714 data pairs of body surface potential (BSP) signal and ischemic volume were extracted across the 8 experiments. The data pairs were then randomly distributed into a 90%-10% training-validation split, where the training dataset had 71,743 data pairs and the validation dataset had 7,971 data pairs. The same training-validation split pattern was applied to all subsequent analyses.

Machine Learning Architecture and Training for ECG:

The ischemic volumes were predicted using a regression training with a custom convolutional neural network similar to that described by Bergquist et al.[14,15] The network architecture was modified to have one spatial and one temporal resolution, with a 7×1 filter and an 1×5 filter, respectively. Additionally, a softplus was added to the output to ensure the target values stay positive.[14,15] The networks were trained with the mean squared error loss between the network predictions and the target ischemic volumes using the body surface ECGs as inputs, and the Adam optimizer was used to fine-tune the weights. Each network was trained for 50 iterations, where the training and validation R² values were monitored to evaluate network performance. Five replicates of training and validation steps were completed to minimize the effects of initialization. All machine learning was implemented using the Pytorch library.[16]

Analysis metrics:

R² value (coefficient of determination) was used to evaluate network performance for both the training and testing sets. In addition, the root mean square error (RMSE) was calculated for each of the 5 networks. The validation dataset was also examined and categorized based on the size of the ischemic tissue. An ischemic region above 10 cm³ was considered a large ischemic region (10.39%, n = 828), whereas a region below 10 cm³ was considered a small or no ischemic region (89.61%, n = 7,143).

3. Results

Five networks were trained and evaluated on the ischemic volume dataset. Across all networks, the peak R² value for the validation dataset was found to be 0.937, and the average R² value was 0.932 ± 0.004 (Table 1).

Table 1.

The R² and the RMSE values from the validation of the 5 ML networks

Network #	Validation R2	RMSE (cm3)
1	0.936	1.084
2	0.931	1.129
3	0.937	1.079
4	0.926	1.171
5	0.931	1.130
Mean	0.932	1.119
±Standard Deviation	±0.004	±0.034

The RMSE for each network was also calculated, with the smallest error being 1.079 cm³ and the average being 1.119 ± 0.034 cm³ (Table 1). In general, as the R² value increases, the RMSE decreases in line with the improved accuracy. The error can be more directly represented in a regression curve (Figure 2).

Figure 2.

The sample regression curve of the network performance on validation data. The predicted volumes by the ML network were compared to the real volumes (blue circles, units of cm³). A linear line of best fit was also visualized (orange). For demonstration purposes, the best-performing network #3 was used for the error evaluation here.

Among the large ischemic region data, the average RMSE across the 5 networks was found to be 1.900 ± 0.067 cm³, which represented a 70% increase compared to the overall RMSE average. On the other hand, for data with small or no ischemic regions, the average RMSE was found to be 0.989 ± 0.035 cm³, a 12% decrease compared to the overall RMSE average.

4. Discussion and Conclusions

In this study, 5 ML networks were trained on the same regression task to predict the ischemic tissue volumes on the simultaneously recorded body surface ECGs. The best-performing network achieved a validation R² value of 0.937, and all 5 networks achieved a validation R² value over 0.925. The average RMSE was 1.119 cm³ across all networks, and no networks have an RMSE exceeding 1.180 cm³. The high R² values and the low errors show that the networks have robust performances when predicting ischemic volumes directly from ECGs, a feature that traditional ECG interpretation does not allow.

When comparing the prediction accuracy across different ischemic sizes, the networks generally performed better for small to no ischemic regions, according to the decrease in the average RMSE. Such results may indicate that as the ischemic region size grows, the complexity of the signal and difficulty in predicting its volume also grow. The geography of the ischemic regions might also contribute to this behavior. Particularly, the ischemic regions usually localize to one territory in the heart, whereas the body surface ECG data records over all territories. The training dataset could also have more low ischemic volume data points, leading to the bias in the performance.

The performance of the networks in ischemic volume prediction suggests that the spatiotemporal design of the network is advantageous in learning the features contained in a multi-electrode ECG recording. Specifically, the design can incorporate the spatial correlations of the signals, due to their proximity in location, in addition to the temporal features of the ECG waveform. The discovery leads the way for an array of future applications of this network design. In addition to independent ECG analysis, the ML tools presented in this study can also be combined with electrocardiographic imaging (ECGI) techniques for localizing myocardial ischemia.[17] ECGI techniques can leverage knowledge of the torso anatomy, electrode locations, and electrical tissue conductivities, thus providing a higher spatial resolution in estimating the bioelectric sources in the heart than traditional 12-Lead ECGs.[17] Such combined physics-based (ECGI) and data-driven (ML) modeling may provide results superior to either alone.

One of the most substantial obstacles for ML network training is the lack of high-quality data. This lack is a problem especially when targeting human diseases, where the availability of detailed datasets is often limited. In this study, we leveraged a large animal model dataset to train ML networks. In the future, other ML approaches could benefit from training with animal datasets, and then possibly transitioning or transfer-learning to a human-specific yet limited dataset.

Future studies should include the application of these animal dataset-trained ML networks to various downstream tasks, such as other predictions within the animal dataset and possibly transitions to human datasets. Explainability analysis would help elucidate the mechanisms behind the ML-ECG analysis, given that myocardial ischemia produces a readily recognizable ECG signature. Furthermore, the datasets in this study included all the recordings across multiple varied ischemic interventions to represent a more diverse and representative cohort. However, the data can be further divided into preischemic and ischemic recordings, which may allow for more complicated ML tasks such as predicting ischemic response or outcome using preischemic signals. Such a risk-stratification ML tool that uses only preischemic ECG recordings would be helpful for clinical management. Lastly, the scope of this study was limited to predicting myocardial ischemic size only. Future applications of the same approach could expand to other clinical metrics of interest and create more useful tools for diagnosing and monitoring cardiac diseases.