Deep learning on electrocardiogram waveforms to stratify risk of obstructive stable coronary artery disease

Rishi K Trivedi, I-Min Chiu, J. Weston Hughes, Albert J. Rogers, David OuyangEuropean 2025 · DOI

TL;DR

This paper introduces a deep learning (DL) model that uses electrocardiogram (ECG) waveforms and clinical data to predict obstructive coronary artery disease (oCAD). The multi-modal DL approach significantly outperforms traditional risk scores, offering a more accurate and potentially earlier method for identifying patients who need further cardiac evaluation.

What Problem Does This Paper Address?

Coronary artery disease (CAD) remains a leading cause of death, with a growing burden of chronic CAD (CCD). Current methods for assessing the risk of obstructive CAD (oCAD) – the narrowing of coronary arteries that can lead to heart attacks – rely on pre-test probability (PTP) scores. These scores, based on factors like age, sex, medical history, and chest pain symptoms, have limitations: they tend to overestimate risk in some populations (leading to unnecessary testing) and underestimate it in others (like women and minorities), potentially delaying diagnosis and treatment. The authors aim to develop a more accurate and equitable risk stratification tool for oCAD.

Methods

The researchers developed three deep learning models:

  • DL-ECG: Trained solely on raw 12-lead ECG waveform data.
  • DL-Clinical: Trained on traditional clinical features used in existing risk scores.
  • DL-MM (Deep Learning Multi-Modal): A combined model that integrates both ECG waveform data and clinical features.

These DL models were trained using neural network architectures, including convolutional neural networks (CNNs) for ECG data and feedforward neural networks (FNNs) for clinical data. The performance of these models was compared against the CAD2 risk score, a widely used PTP estimation tool. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), a standard metric for classification accuracy.

Datasets

The study utilized two main datasets:

  • Cedars-Sinai Medical Center (CSMC) Cohort: This was the primary dataset for model development and internal validation. It included 8,361 patients undergoing coronary angiography, with 40,602 ECGs collected within five years prior to the procedure.
  • Stanford HealthCare (SHC) Cohort: This dataset was used for external validation. It comprised 17,012 patients with 70,755 ECGs collected between August 2005 and May 2018.

The primary outcome for labeling was the presence of obstructive CAD, defined by whether percutaneous coronary intervention (PCI) was performed during angiography.

Key Insights and Ideas

  • ECG Waveforms Hold Significant Predictive Power: Even when used alone, DL analysis of raw ECG waveforms can predict oCAD with accuracy comparable to models relying solely on clinical features. This suggests that subtle patterns in ECGs, often imperceptible to human clinicians, contain valuable diagnostic information.
  • Multi-Modal Approach is Superior: Combining ECG waveform data with clinical features in a DL model (DL-MM) significantly enhances predictive performance beyond either modality alone or traditional risk scores.
  • Potential for Improved Equity: The DL-MM model showed more consistent performance across various demographic and clinical subgroups compared to traditional risk scores, which have known biases. This suggests DL could lead to more equitable risk assessment.
  • Long-Term Predictive Value: The DL-MM model demonstrated predictive capabilities for oCAD even when ECGs were taken up to five years before angiography, hinting at its ability to detect underlying sub-clinical atherosclerotic changes.

Key Results

  • Internal Validation (CSMC Hold-out Test Set):
    • CAD2 (traditional risk score): AUC 0.733
    • DL-ECG (ECG only): AUC 0.741
    • DL-Clinical (clinical features only): AUC 0.762
    • DL-MM (multi-modal): AUC 0.807 (statistically superior to CAD2)
  • External Validation (SHC Cohort):
    • CAD2: AUC 0.715
    • DL-MM: AUC 0.716 (similar performance to CAD2)
    • DL-ECG: AUC 0.633 (significantly lower than DL-MM and CAD2)

The DL-MM model consistently outperformed the CAD2 model in the internal validation, showing a statistically significant improvement. While its performance slightly decreased in the external validation, it remained comparable to the CAD2 score.

Limitations and Future Work

  • Retrospective Nature and Cohort Differences: The study was retrospective, and differences in patient demographics, co-morbidities, and oCAD prevalence between the CSMC and SHC cohorts likely contributed to the performance drop in external validation.
  • Definition of oCAD: Using PCI as a surrogate for oCAD is partially subjective and may not capture all cases of significant CAD, especially those managed with surgery or not amenable to PCI.
  • Model Explainability: While SHAP analysis identified ECG waveform analysis as the most influential feature, the specific ECG components driving the predictions remain difficult to interpret with current DL explainability methods. This is a barrier to clinical implementation.
  • Generalizability: Further research is needed to evaluate the generalizability of the DL model across diverse populations and healthcare settings.
  • Clinical Utility and Safety: Prospective studies are required to determine if incorporating DL methods into ECG analysis improves oCAD diagnosis, patient outcomes, and clinical workflow efficiency. The authors suggest that future work should explore alternative labels for oCAD (e.g., anatomical stenosis, functional ischemia) and investigate the model's performance in low-resource settings.

Summary updated

Abstract

A multi-modality DL model utilizing ECG waveforms and clinical risk factors can improve prediction of oCAD in CCD compared with risk-factor based models. Prospective research is warranted to determine whether incorporating DL methods in ECG analysis improves diagnosis of oCAD and outcomes in CCD.
BibTeX
@inproceedings{trivedi2025deep,
  author = {Rishi K Trivedi and I-Min Chiu and J. Weston Hughes and Albert J. Rogers and David Ouyang},
  title = {{Deep learning on electrocardiogram waveforms to stratify risk of obstructive stable coronary artery disease}},
  booktitle = {European},
  year = {2025},
  doi = {10.1093/ehjdh/ztaf020},
}
Figure