Machine Learning Modeling of 30-Day Outcomes in Low-Risk Patients Suspected of Myocardial Infarction

Document Type

Conference Proceeding

Publication Date

5-13-2025

Publication Title

Acad Emerg Med

Keywords

troponin, troponin I, acute heart infarction, adult, algorithm, benchmarking, cellular distribution, cohort analysis, conference abstract, controlled study, cross validation, diagnosis, emergency ward, female, heart infarction, human, incidence, limit of quantitation, low risk patient, machine learning, major clinical study, male, oxygen saturation, random forest, secondary analysis

Abstract

Background and Objectives: Existing machine learning (ML) models for acute myocardial infarction (AMI) have primarily been trained using data with a high incidence of AMI. While these models are helpful, their outcomes are self-evident and fail to address what emergency department (ED) clinicians are most worried about-is my patient without clear AMI in the ED at risk of adverse events at 30 days? Our objective was to evaluate ML models of 30-day death and AMI in low-risk patients. Methods: A secondary analysis of a large, prospective implementation trial of high-sensitivity cardiac troponin I (hs-cTnI) across nine diverse EDs. Patients with cardiopulmonary symptoms necessitating evaluation with cTn testing were eligible. We excluded those with trauma, age < 18 years, or any hs-cTnI > the 99th percentile. We trained three ML algorithms (logistic regression, random forests, and gradient boosting) and validated them through 5-fold cross-validation. The primary outcome was death or AMI within 30 days. To correct class imbalance, we used the oversampling algorithm SMOTE (Synthetic Minority Oversampling Technique) and grid search optimized hyperparameters. We evaluated model performance with the AUC, accuracy, recall (true positive rate), precision, and F1-score. We selected the most effective model based on AUC. We used feature importance analysis to identify significant predictors. Results: The analysis included 28,805 patients, of whom 116 (0.4%) had death/MI within 30 days. There were 16,713 (58%) females, 9.390 (33%) Black patients, and the mean age was 58.7 ± 17.9 years. The random forest model achieved the highest AUC (0.62) and F1-score (0.48), higher than the other algorithms on all metrics (logistic regression AUC 0.45 and F1-score 0.44, gradient boosting AUC 0.50 and F1-score 0.47). Recall was 0.55 with a precision of 0.50 and 90% accuracy. Major 30-day death or MI predictors in order of weight in the model were elevated troponin (between the limit of quantification and 99th percentile), race, oxygen saturation, age, and red cell distribution width. Class imbalance mitigation strategies improved model performance but failed to improve above an AUC of 0.62. Conclusion: Random Forest performed best in modeling 30-day events in low-risk patients at risk of AMI. Nonetheless, the low-event rate substantially limits the added value of ML models for this population.

Volume

32

Issue

S1

First Page

20

Find It @ Sladen

Share

COinS