Development and Validation of Artificial Intelligence Prediction of Epicardial Coronary Artery Spasm in Patients Without Obstructive Coronary Artery Disease

Figures

• 

  Figure 1. ECG CNN–LSTM classification pipeline for CAS versus control. (A) Dataset composition and stratified training, validation, and test split of CAS and control ECG samples. (B) Preprocessing workflow converting raw ECG images into fixed-length four-lead numerical signals, followed by tensor stacking and per-lead z-score normalization. (C) Hybrid CNN–LSTM architecture, in which Conv1D blocks extract local temporal ECG features and LSTM layers capture sequential dependencies before sigmoid classification. (D) Training strategy, including optimizer, loss, metrics, class weighting, callbacks, evaluation outputs, and saved model files. BN, Batch Normalization (normalizes layer inputs to stabilize and accelerate training); Conv1D, 1D Convolution (a one-dimensional convolutional layer used to extract features from time-series signals); CSV, Comma-Separated Values (a tabular data file format where values are separated by commas); ECG, electrocardiogram; ReLU, Rectified Linear Unit (a common activation function, f(x) = max(0, x)); Val, Validation.

• 

  Figure 2. The relative importance of predictors included in the initial random forest model in the derivation cohort.

• 

  Figure 3. Discriminative performance of the initial and final random forest models in the derivation cohort (A) and performance of the final random forest model in the validation cohort (B). AUC, area under the curve; CI, confidence interval.

• 

  Figure 4. Partial dependence plots derived from the final random forest model in the derivation cohort, including diastolic blood pressure (A), systolic blood pressure (B), age (C), body surface area (D), hemoglobin (E), smoking (F), heart rate (G), sex (H) and estimated glomerular filtration rate (I). The light red–shaded area indicates the range along the X-axis where the probability of CAS is relatively higher. CAS, coronary artery spasm.

• 

  Figure 5. The probability of CAS stratified by the number of predictors in the (A) derivation and (B) validation cohorts. Variables based on the thresholds from the partial dependence plots were determined from the partial dependence plots. CAS, coronary artery spasm; CI, confidence interval; OR, odds ratio.