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

Figures

• 

  Figure 1. Consort diagram. Consort diagram depicting the internal and external validation cohorts. Exclusions were made for subjects where indication for coronary angiography was an acute coronary syndrome and if the subject did not have an ECG within 5 years of coronary angiography. These studies were paired with all ECGs obtained within 5 years of coronary angiography. Label for presence of obstructive coronary artery disease was whether percutaneous coronary intervention was performed.

• 

  Figure 2. AUC of model performance. Visual representation of the area under the receiver-operator curve characteristics including all developed models. The deep learning multi-modal model had a visually distinct performance curve consistent with its statistically significant superior performance as seen in Table 2. AUC, area under the receiver-operator curve; CAD2, coronary artery disease consortium risk estimation tool; DL-clinical, deep learning model trained on clinical features alone; DL-ECG, deep learning model trained on ECG waveforms only; DL-MM, deep learning model trained on ECG waveforms and clinical features.

• 

  Figure 3. Model performance across time for detection of obstructive coronary artery disease. Changes in model performance based on timing of ECG prior to coronary angiogram. The deep learning multi-modal model had superior performance compared with the CAD2 model, which was preserved regardless of whether the ECG was obtained within 1 year of the angiogram or from 1 to 5 years prior to angiography. AUC, area under the receiver-operator curve; CAD2, coronary artery disease consortium risk estimation tool; DL-clinical, deep learning model trained on clinical features alone; DL-ECG, deep learning model trained on ECG waveforms only; DL-MM, deep learning model trained on ECG waveforms and clinical features; PCI, percutaneous coronary intervention.

• 

  Figure 4. Multi-modal model performance stratified by subgroup. Forest plot depicting performance of the multi-modal model stratified by subgroups. The deep learning multi-modal model had mostly consistent performance regardless of subgroup. The model had slightly reduced performance in younger subjects, hypertensive patients, and diabetic patients; the model had slightly higher performance in current smokers. AUC, area under the receiver-operator curve; CAD2, coronary artery disease consortium risk estimation tool; CI, confidence interval; LVH, left ventricular hypertrophy.

• 

  Figure 5. Shapley additive explanations (SHAP) analysis of multi-modal model. SHAP analysis of the multi-modal model depicting, which features included in the multi-modal model had the most impact on model output. The main feature that impacted multi-modal model output was ECG waveform analysis followed by age and chest pain type with limited effect on model output based on the remaining clinical variables. DM, diabetes mellitus; HTN, hypertension; SHAP, Shapley additive explanations.