Patient Stratification for Improving Acute Chest Pain Management and Mitigating Emergency Department Crowding: Machine Learning Model Development and Validation.
Hsieh C, Chu S, et al. • JMIR medical informatics • 2026
ANN-based classifiers can effectively support clinical risk stratification and disposition decision-making in ACP care, but should be interpreted as a clinical decision-support aid rather than an independent rule-out strategy because even the sensitivity-prioritized variant falls short of the stringent sensitivity threshold (>0.99) typically required for a standalone ED rule-out tool.
Key Findings
Background
Among ED patients presenting with chest pain suggestive of cardiac origin, fewer than 10% are ultimately diagnosed with acute coronary syndrome.
Acute chest pain accounts for approximately 8% of all ED visits.
Of 17,935 visits included in the study, only 1,209 (6.7%) were classified as ACS (subgroup GA).
The remaining visits comprised 3,873 critical patients without ACS (GB1) and 12,853 low-risk patients (GB2).
Data were drawn from a single tertiary teaching hospital between January 2016 and December 2022.
Results
ANN models demonstrated strong testing AUROC performance for classifying acute chest pain patients into three clinical subgroups.
AUROC (95% CI) for ACS (GA) classification was 0.942 (0.920–0.965).
AUROC for critical non-ACS (GB1) classification was 0.824 (0.808–0.841).
AUROC for low-risk (GB2) classification was 0.893 (0.884–0.902).
Models were trained using 24 feature variables: 2 demographics, 6 vital signs, 4 blood test results, and 12 medical history items.
Training and internal validation used a 5-fold cross-validation protocol on 2016–2020 data; testing was performed on hold-out data from 2021–2022.
Results
The balanced multiclassification model ANN-S3 achieved high ACS sensitivity and strong low-risk classification performance.
ACS sensitivity was 0.941 (95% CI 0.908–0.973).
Low-risk positive predictive value was 0.911 (95% CI 0.900–0.921).
Low-risk sensitivity was 0.837 (95% CI 0.824–0.850).
ANN-S3 was described as achieving 'balanced performance' across all three clinical classes.
Results
The sensitivity-prioritized variant ANN-S3-L increased ACS detection but introduced a safety-efficiency trade-off.
ACS sensitivity increased to 0.966 (95% CI 0.940–0.991) with ANN-S3-L.
Negative predictive value reached 0.998 (95% CI 0.996–0.999).
Gains in ACS sensitivity came 'at the cost of lower specificity and reduced low-risk sensitivity.'
Despite this improvement, the 0.966 sensitivity still falls short of the >0.99 threshold typically required for a standalone ED rule-out tool.
Discussion
The AI models are intended as clinical decision-support aids rather than independent rule-out strategies.
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.'
The system was designed to integrate with clinical workflows to reduce ED length of stay and alleviate crowding.
Models relied on a single high-sensitivity cardiac troponin T test result and are applicable for patients presenting ≥3 hours after symptom onset.
Feature selection methods and SHAP (Shapley Additive Explanations) value analysis were applied to identify features driving predictive power.
Methods
After data exclusions for missing or unsuitable records, 17,935 ED visits were included in the study cohort.
Exclusions were applied for missing triage data, incomplete medical histories, or unsuitable dispositions.
The dataset spanned January 2016 to December 2022 from a single tertiary teaching hospital (National Cheng Kung University Hospital, Tainan, Taiwan).
The class distribution was highly imbalanced: 1,209 GA (6.7%), 3,873 GB1 (21.6%), and 12,853 GB2 (71.7%).
The study was approved by the institutional review board (A-ER-111-199).
What This Means
This research suggests that artificial intelligence can help emergency department doctors better identify which chest pain patients are having a heart attack (acute coronary syndrome, or ACS), which are critically ill from other causes, and which are low-risk and could be safely discharged sooner. The researchers built and tested AI models using data from nearly 18,000 emergency visits at a hospital in Taiwan over seven years. The models used 24 pieces of information routinely collected at triage — including vital signs, a single blood test for heart damage (troponin T), and medical history — to sort patients into three groups. The best-performing balanced model correctly identified about 94% of ACS cases, while a safety-focused version pushed that rate to nearly 97% with an extremely high negative predictive value of 99.8%, meaning it very rarely missed a true ACS case.
However, the research also found an important limitation: even the most sensitive AI version did not quite reach the 99%+ sensitivity threshold that clinical guidelines typically require before a tool can be used on its own to 'rule out' a heart attack. This means the AI should be used alongside — not instead of — a doctor's judgment. The models are designed specifically for patients who arrive at the emergency department at least three hours after their chest pain started, which is when a single troponin test becomes most reliable.
This research suggests that integrating such AI tools into emergency department workflows could help reduce the time low-risk patients spend in the ED, ease crowding, and improve overall efficiency — all while maintaining strong safety standards for catching serious heart events. The practical implication is that AI could flag low-risk patients for faster discharge and alert clinicians to high-risk patients more quickly, though human oversight remains essential.
Hsieh C, Chu S, Lee J, Lin C, Kao C. (2026). Patient Stratification for Improving Acute Chest Pain Management and Mitigating Emergency Department Crowding: Machine Learning Model Development and Validation.. JMIR medical informatics. https://doi.org/10.2196/83099