1342288 - Machine Learning Decision Tree Models Using Cardiac MRI May Better Predict Etiology of Myocardial Infarction and Injury Compared to Regression Models in Patients with Chest Pain and Elevated Troponin

Background: Myocardial infarction (MI) is the major clinical presentation of coronary artery disease (CAD) in the United States. Recently, the incidence of type 2 myocardial infarction (non-plaque rupture MI) and non-MI myocardial injury has increased compared to type 1 MI (plaque rupture MI); non-plaque rupture can often pose as a clinical diagnostic challenge. Cardiac MRI (CMR) may provide diagnostic clarity for the mechanism of ischemia, and novel machine learning techniques may further augment differentiation between plaque rupture MI versus non-plaque rupture and myocardial injury events. 

Methods: We conducted a sub-analysis of the Cardiac Magnetic Resonance Imaging Strategy for the Management of Patients With Acute Chest Pain and Detectable to Elevated Troponin trial (CMR-IMPACT; NCT01931852). CMR parameters, including native T1, T2, and extracellular volume fraction (ECV), were measured in 103 adults who presented to the emergency department with chest pain and detectable troponin-I and underwent CMR < 36 hours after presentation. Native T1, T2, and ECV were measured using short-axis maps of the mid left ventricle according to the standard 16-segment model. Multivariable regression analysis for prediction of MI/injury type was performed to assess the effects of CMR parameters and other baseline factors, which included age, race, sex, and initial presenting troponin-I level. A second prediction model using a machine learning decision tree classifier was performed using all covariates; a medium tree model with a 10-fold k-fold partition cross-validation scheme was implemented. Feature ranking was performed to determine which covariates had significant predictive power in the model. Covariates that had low predictive power were eliminated from the final decision tree model. Regression analysis was performed using STATA version 14.1 (StataCorp, College Station, TX), and machine learning modeling was performed using MATLAB version R2021b (Mathworks Inc., Natick, MA). Area under the receiver-operating characteristic curve (AUROC) was computed for both models.

Results: Baseline characteristics and CMR measurements are shown in Table 1. The regression model showed that ECV and initial troponin were significant predictors for the outcome of non-plaque rupture versus plaque rupture events. In contrast, the final tree model included T1, T2, and ECV as significant predictive features; the true positive rate for predicting non-plaque events was 94.3% (Figure 1). AUROC for the regression and decision tree models were 0.87 and 0.94, respectively.

Conclusion: CMR can distinguish between plaque rupture and non-plaque rupture MI and injury in patients presenting with chest pain and elevated troponins. Furthermore, machine learning tree classifiers may yield higher accuracy compared to regression models in predicting ischemic etiology.