Background: Chest pain is the second leading reason for emergency department (ED) visits and is commonly identified as a leading driver of low-value health care. Accurate identification of patients at low risk of major adverse cardiac events (MACE) is important to improve resource allocation and reduce over-treatment.
Objectives: We sought to assess machine learning (ML) methods and electronic health record (EHR) covariate collection for MACE prediction. We aimed to maximize the pool of low-risk patients that are accurately predicted to have less than 0.5% MACE risk and may be eligible for reduced testing.
Population Studied: 116,764 adult patients presenting with chest pain in the ED and evaluated for potential acute coronary syndrome (ACS). 60-day MACE rate was 1.9%.
Methods: We evaluated ML algorithms (lasso, splines, random forest, extreme gradient boosting, Bayesian additive regression trees) and SuperLearner stacked ensembling. We tuned ML hyperparameters through nested ensembling, and imputed missing values with generalized low-rank models (GLRM). We benchmarked performance to key biomarkers, validated clinical risk scores, decision trees, and logistic regression. We explained the models through variable importance ranking and accumulated local effect visualization.
Results: The best discrimination (area under the precision-recall [PR-AUC] and receiver operating characteristic [ROC-AUC] curves) was provided by SuperLearner ensembling (0.148, 0.867), followed by random forest (0.146, 0.862). Logistic regression (0.120, 0.842) and decision trees (0.094, 0.805) exhibited worse discrimination, as did risk scores [HEART (0.064, 0.765), EDACS (0.046, 0.733)] and biomarkers [serum troponin level (0.064, 0.708), electrocardiography (0.047, 0.686)]. The ensemble's risk estimates were miscalibrated by 0.2 percentage points. The ensemble accurately identified 50% of patients to be below a 0.5% 60-day MACE risk threshold. The most important predictors were age, peak troponin, HEART score, EDACS score, and electrocardiogram. GLRM imputation achieved 90% reduction in root mean-squared error compared to median-mode imputation.
Conclusion: Use of ML algorithms, combined with broad predictor sets, improved MACE risk prediction compared to simpler alternatives, while providing calibrated predictions and interpretability. Standard risk scores may neglect important health information available in other characteristics and combined in nuanced ways via ML.
Acronyms and Opportunities for Improving Deep Nets
