Artificial intelligence-enhanced electrocardiography improves the detection of coronary artery disease

Figures

• 

  Fig. 1. Patient flow diagram and examples that support the performance of the AI-enhanced ECG approach. (A) ECGs are included in this retrospective study under the following criteria, that they are the latest ECG records and were obtained no more than 1 month before a patient received coronary angiography (CAG). (B) The training cohort is composed of four groups. Group 0: normal ECG (sinus rhythm only) obtained from subjects at a young age (18–39 years old); this group is to balance the dominant number of CAD patients. Group 1: normal ECG with a description of sinus rhythm only. Group 2: abnormal ECG without an ischemic description. Group 3: abnormal ECG with an ischemic description (ischemic descriptions include the following key words: MI, infarct, infarction, ischemia and subendocardial injury).

• 

  Fig. 2. Performance of the AI-enhanced ECG when detecting CAD. (A) Flow chart of ML, including data preparation, signal pre-processing and feature engineering using training and validation cohorts. The network accepts raw ECG xml data as an input and outputs a prediction of CAD probability. Performance of the AI-algorithm is carried out by 5-fold cross validation and evaluated using the results of the area under curve (AUC) under the Receiver Operator Characteristic Curve (ROC) approach, accuracy rate (ACC), sensitivity, specificity, positive (PPV) and negative (NPV) predictive values. (B) The AI-algorithm enhances the performance of the ECGs compared with conventional ECGs in the training cohort. F1 score is calculated as the harmonic mean of the precision and recall scores. (C) The prediction rates of the AI-enhanced ECG algorithm in subjects with normal sinus rhythm (Groups 0 + 1), Group 2 and Group 3 from the training cohort. (D) ROC curves calculated for the detection of CAD in all subjects and subjects with a normal sinus rhythm (Group 0 + Group 1). Green curve is the training cohort, while orange curve is the validation cohort. The AUC values for the training and validation…

• 

  Fig. 3. Sex differences, examples of the top ranked ECG features and the composition of the probability of CAD prediction in CAD and non-CAD subjects. (A-C) Examples presented are the three top ranked features (rank #1, #3 and #6) of males. (D-F) Examples presented are the three top ranked features (rank #1, #2 and #3) of females. Box plots of the feature with statistically significant differences between CAD and non-CAD patients are shown. The probability of CAD prediction is calculated by adding the Shapley additive explanations (SHAP) values of the top nine features and the remaining 552 features. A positive SHAP value (red bar) means a contribution to the probability of CAD and a negative SHAP value (blue bar) means the opposite. Each individual has a personal list of feature ranked by SHAP framework. * ** *p < 0.0001.

• 

  Fig. 4. Comparison of Diagnostic Modalities for Coronary Artery Disease. (A) An example of a 61-year-old male CAD patient who has a normal ECG and a negative result by Thallium scanning. However, his CAG result revealed that he has a significant infarction in the LAD and LCX coronary arteries. Interestingly, the AI-enhanced ECG algorithm predicts the probability of CAD to be 0.75 for this subject. (B) Another example of a 60-year-old male non-CAD patient, who has a normal ECG and a negative Thallium scan; he has normal coronary arteries as revealed by CAG. Notably, the AI-enhanced ECG algorithm predicts a low probability of CAD, namely 0.27, for this subject. (C) Our AI-enhanced ECG algorithm should be able to provide an alternative approach when identifying high-risk subjects with potential CAD. For patients with chest pain, initial examination by resting 12-lead ECG, followed by nuclear medicine examination, namely myocardial perfusion scintigraphy (MPS)-Thallium scanning or cardiac computerized tomography (CT), then invasive coronary angiography (CAG), is recommended currently. Our AI-enhanced ECG algorithm has a performance comparable with MPS-Thallium scanning and thus it may serve…