Machine learning-based prediction of acute coronary syndrome using only the pre-hospital 12-lead electrocardiogram

Figures

• 

  Fig. 1. Stages of dataset derivation and preparation prior to developing the ML classifiers. We used all available ECG features (k = 554), selected ECG features (k = 65), and selected and relabeled ECG features (k = 65 + L) to train and test our machine learning (ML) classifiers: logistic regression (LR), gradient boosting machine (GBM), and artificial neural networks (ANN). Cohort 1 was used for training and Cohort 2 was used for independent testing. The primary study outcome was acute coronary syndrome (ACS).

• 

  Fig. 2. Classification performance using machine learning classifiers. This figure shows the ROC curves of logistic regression (LR), gradient boosting machine (GBM), and artificial neural network (ANN) classifiers using a all available ECG features (k = 554), b selected ECG features (k = 65), and c selected and relabeled ECG features (k = 65 + L). Cohort 1 was used for training with tenfold cross-validation, the figure shows mean ROC curve with ±2 standard errors. Cohort 2 was used for independent testing using the algorithm trained on Cohort 1.

• 

  Fig. 3. Classification performance of final model on testing set (n= 499). This figure compares the area under ROC curve (95% confidence interval) between our machine learning (ML) fusion model against experienced clinicians and against rule-based commercial interpretation software for detecting a any acute coronary syndrome (ACS) event, and b non-ST elevation acute coronary syndrome events (NSTE-ACS). ***p < 0.001 using two-sided DeLong’s nonparametric approach.

• 

  Fig. 4. Scatterplot matrix of some selected features using recursive feature elimination. This figure shows selected ECG features with the two-dimensional display scatterplot matrices. These plots show how linear correlations fail to separate patients with or without acute coronary syndrome (ACS), which explains why nonlinear classifiers were computationally favored over linear classifiers in our study.