Key result
LightGBM distinguishes ACS from non-ACS in chest pain patients with ~0.98 AUC.
Why the study?
Rapid identification of ACS among patients presenting with chest pain remains challenging because several cardiovascular and noncardiovascular conditions have similar initial presentations.
Does explainable machine learning using routinely collected chest pain center data improve early identification of acute coronary syndrome in patients presenting with chest pain?
Observational (n=68,224)
Yes
Does explainable machine learning using routinely collected chest pain center data improve early identification of acute coronary syndrome in patients presenting with chest pain?
Answers are grounded in this paper and cite their sources.
Explainable machine learning models, particularly LightGBM, using early structured clinical and laboratory data can accurately distinguish acute coronary syndrome from non-ACS conditions in patients presenting with chest pain.
Background Rapid identification of acute coronary syndrome (ACS) among patients presenting with chest pain remains challenging because several cardiovascular and noncardiovascular conditions may have similar initial presentations. We aimed to develop and internally validate explainable machine learning models for the early identification of ACS using routinely collected chest pain center data. Methods This retrospective diagnostic prediction model development and internal validation study used data from the Chinese Cardiovascular Association (CCA) Database-Chest Pain Center, collected between January 2015 and December 2023. Patients with complete core clinical variables and at least one valid laboratory result were included. The cohort was randomly divided into training and held-out test sets at a ratio of 8:2 using stratified sampling. Six models were developed, including ElasticNet, random forest, gradient boosting, XGBoost, LightGBM, and CatBoost. Hyperparameters were optimized using nested five-fold cross validation and random search. Model performance was assessed using discrimination, calibration, classification metrics, and decision curve analysis. SHapley Additive exPlanations (SHAP) were used to interpret the final model. Results A total of 68,224 patients were included, of whom 48,474 had ACS and 19,750 had non-ACS conditions. The median age was 64.0 years, and 63.9% were men. All models showed high discrimination in the held-out test set, with areas under the receiver operating characteristic curve ranging from 0.9528 to 0.9818. LightGBM achieved the highest area under the curve of 0.9818 (95% CI, 0.9800 to 0.9837) and balanced accuracy of 0.9497 (95% CI, 0.9461 to 0.9531). At the selected classification threshold, its sensitivity was 0.9196, specificity was 0.9797, and F1 score was 0.9540. Calibration was good, and decision curve analysis (DCA) indicated potential theoretical net benefit within the internal validation cohort; these DCA findings require external validation. SHAP analysis showed that predictions were jointly influenced by cardiovascular risk factors, previous coronary heart disease, cardiac troponin, D-dimer, other laboratory markers, vital signs, and chest pain characteristics. Conclusions Machine learning models based on structured clinical information available early during presentation accurately distinguished ACS from non-ACS conditions in a large real-world chest pain center cohort. LightGBM showed the strongest overall performance and provided clinically interpretable predictions. External validation and prospective evaluation are required before routine clinical implementation, and the DCA findings should be regarded as theoretical estimates.
No takes yet. Share an insight, caveat, or question.
Yin et al. (2026) conducted an observational in Acute coronary syndrome (ACS) among patients presenting with chest pain (n=68,224). Explainable machine learning models (LightGBM, ElasticNet, random forest, gradient boosting, XGBoost, CatBoost) was evaluated on Early identification of acute coronary syndrome (ACS) vs non-ACS conditions (AUC 0.9818, 95% CI 0.9800-0.9837). LightGBM accurately distinguished ACS from non-ACS in chest pain patients, achieving an AUC of 0.9818 and balanced accuracy of 0.9497.
Synapse has enriched 10 related papers. Consider them for comparative context: