Multisite derivation of a machine learning algorithm using high sensitivity troponin to predict major adverse cardiac...

To develop a machine learning (ML) algorithm to stratify risk for major adverse cardiac events (MACE) within 30 days in emergency department (ED) patients undergoing troponin testing.

Retrospective cohort analysis using extreme gradient boosting (XGBoost), a tree-based ensemble machine learning algorithm.

Twenty U.S. hospitals.

Patients aged ≥22 years who underwent high-sensitivity troponin-I (Beckman Coulter Access high-sensitivity troponin-I, hs-cTnI) testing between October 2019 and December 2020.

We evaluated ML model performance for predicting 30-day MACE using negative predictive value (NPV), sensitivity, and specificity. The model used only objective EHR data.

Out of 95,093 ED visits, 91,278 met inclusion criteria. The ML model generated predictions at three clinical timepoints based on troponin availability: initial (all patients), second (subset with serial testing), and final (patient's last result, with an area under the receiver operating characteristic curve (AUROC) of 0.90 (95% CI 0.89-0.90) for the initial prediction and 0.91 (95% CI 0.90-0.91) at the final troponin result. It identified 53.2% of patients as low risk with an NPV of 99.35% (95% CI 99.17% to 99.49%). The model showed strong calibration and discrimination, particularly in its ability to safely increase the proportion of patients classified as low risk for discharge.

This study demonstrates feasibility of automated machine learning using objective EHR data to predict 30-day MACE among ED patients undergoing troponin testing. This approach minimizes subjective interpretations and warrants prospective validation to assess potential for improving ED efficiency and patient safety.