Patient Stratification for Improving Acute Chest Pain Management and Mitigating Emergency Department Crowding: Machine Learning Model Development and Validation.

Background: Acute chest pain (ACP) is one of the most common chief complaints in the emergency department (ED), accounting for approximately 8% of all ED visits. However, among patients presenting with chest pain suggestive of cardiac origin, fewer than 10% are ultimately diagnosed with acute coronary syndrome (ACS). Objective: AI models were developed to support clinical decision-making for triage level-2 ED patients presenting with ACP. These models aim to accurately detect ACS and reliably identify low-risk patients based on a single high-sensitivity cardiac troponin T test result. By integrating this AI-assisted strategy into clinical workflows, we aim to reduce ED length of stay, alleviate crowding, and improve health care efficiency while maintaining high safety standards. Methods: We conducted a retrospective study using single-center data from a tertiary teaching hospital between January 2016 and December 2022. Models based on artificial neural networks (ANNs) were trained to classify patients with ACP at triage level 2 into 3 clinical classes: critical patients with ACS (subgroup GA), critical patients without ACS (subgroup GB1), and low-risk patients (subgroup GB2). The models were trained and internally validated via a 5-fold cross-validation protocol using data from 2016 to 2020. Model performance was then evaluated on the hold-out testing data (2021-2022) using area under the receiver operating characteristic curve, area under the precision-recall curve, and subgroup-specific metrics. Feature selection methods and the Shapley Additive Explanations value analysis were applied to identify features that drive the model's predictive power. The study was approved by the institutional review board of the National Cheng Kung University Hospital, Tainan, Taiwan (A-ER-111-199). Results: After excluding ED visits with missing triage data, incomplete medical histories, or unsuitable dispositions, 17,935 visits were included, with 1209 GA, 3873 GB1, and 12,853 GB2. Rendering prediction based on 24 feature variables (2 demographics, 6 vital signs, 4 blood test results, and 12 medical history), all ANN models demonstrated strong testing AUROC (95% CI) performance of 0.942 (0.920-0.965) for GA classification, 0.824 (0.808-0.841) for GB1, and 0.893 (0.884-0.902) for GB2. Regarding multiclassification, ANN-S3 achieved balanced performance, with ACS sensitivity of 0.941 (95% CI 0.908-0.973) and low-risk positive predictive value and sensitivity of 0.911 (95% CI 0.9-0.921) and 0.837 (95% CI 0.824-0.85), respectively. A sensitivity-prioritized variant, ANN-S3-L, increased ACS sensitivity to 0.966 (95% CI 0.94-0.991) and negative predictive value to 0.998 (95% CI 0.996-0.999), but at the cost of lower specificity and reduced low-risk sensitivity, indicating a safety-efficiency trade-off. Conclusions: These findings suggest that ANN-based classifiers can effectively support clinical risk stratification and disposition decision-making in ACP care for patients presenting ≥3 hours after symptom onset. However, because even the sensitivity-prioritized ANN-S3-L variant falls short of the stringent sensitivity threshold (>0.99) typically required for a standalone ED rule-out tool, this system should be interpreted as a clinical decision-support aid rather than an independent rule-out strategy.