Predicting triage levels in patients presenting with cardiac-related symptoms: a comparison of supervised machine learning methods

Comprehensive Summary

This study evaluated whether supervised machine learning models can accurately predict Emergency Severity Index (ESI) triage levels for patients presenting with triage levels for patients presenting with cardiac related symptoms. Researchers prospectively collected 1,862 patients’ records from a specialized cardiac ED and extracted 27 clinical features identified through literature review as well as surveys. Five machine learning methods that included, Random Forest (RF), Gradient Boosting (GB), Support Vector Machine (SVM), Logistic Regression (LR), and K-Nearest Neighbors (KNN), were trained to cross validate and were tested on a test set. The RF model achieved the best performance, with 93.57 % accuracy and 0.82 Cohen’s kappa, and F1 score of 0.93. This outperformed GB which had an accuracy of 90.08 %, SVM had the accuracy of 86.6 %, and LR 76.41 %. Feature important analyses using SHAP showed that high-risk cardiac conditions, need for life-saving interventions, chief complaint, and level of consciousness were the strongest predictors. The conclusion was that machine learning models can reliably replicate nurse assigned triage levels and distinguish between high and low acuity patients.

Outcomes and Implications

This work is important because triage inaccuracies directly influence patient wait times, ED overcrowding, and potentially mortality, especially in cardiac patients who comprised the study’s entire cohort. The strong performance of RF and GB suggests these models could be integrated into ED workflows to support rapid acuity classification, especially for the most common triage levels, level 2 and level 3, which accounted for 89 % of the dataset. Clinically, such systems could be embedded into electronic health records to provide real time decision support using the first fifteen minutes of clinical data, which is all the model requires. Although there is a requirement for external validation and live clinical testing are still needed, the findings indicate that implementation could be feasible within the next several years, improving resource allocation and early identification of high risk cardiac patients.