First-line risk stratification with machine learning models facilitates rapid triage for non-ST-elevation myocardial infarction

Figures

• 

  Fig 1. Patient selection criteria and machine learning procedure. (A) Inclusion and exclusion criteria cascade of patients in the dataset. STEMI, ST-segment elevation myocardial infarction; NSTEMI, non-ST-elevation myocardial infarction; hs-cTnT, high-sensitivity cardiac troponin T. CBC, complete blood count. (B) Schematic diagram of the machine learning model construction pipeline. Routine laboratory testing results completed along with the first hs-cTnT were used to construct a feature vector into which demographics including sex, age, BMI, and smoking status were incorporated. At the endpoint, patients diagnosed with AMI are labeled as “1”, while non-AMI status are labeled as “0”. ML models were built with the feature vector where each dimension represents a specific laboratory test. Model performances were evaluated by receiver-operating-curve (ROC) and importance analysis.

• 

  Fig 2. Prediction performance and interpretability analysis of machine learning models. Comparison of the area under receiver-operating-curves (AUROC) (upper left) and average precision of precision-recall curves (APPRC) (upper right) generated from the (A) testing set and (C) independent validation set for the XGBoost (XGB), random forest (RF), and logistic regression (LR) models, and the model using 0 h hs-cTnT alone as an input feature. Solid lines represent median and color zones represent 95% confidence interval. (B) Importance analysis explaining the contribution of laboratory tests to predicting the model for AMI status. Feature importance is computed by Shapley Additive Explanations (SHAP) values. The laboratory tests are arranged along the y-axis based on their mean absolute SHAP values depicted on the x-axis. These values signify the influence of the test results on the XGBoost model predictions. The color of individual test values for each patient corresponds to their respective relative values.

• 

  Fig 3. Comparison of the discriminative capabilities of the machine learning algorithm and the 0 h/1 h algorithm. Classification cascade of patients with suspected non-ST elevation myocardial infarction using the (A) 0 h/1 h algorithm alone or (B) in combination with a machine learning algorithm. PPV, positive predictive value; NPV, negative predictive value.

• 

  Fig 4. Visual summary delineates the diagnostic cascade incorporating the machine learning algorithm. The combined results demonstrate clinically applicable approaches to rapid risk assessment with this machine learning algorithm for patients presenting with suspected non-ST-elevation myocardial infarction to the emergency department. hs-cTn, high-sensitivity cardiac troponin.