Machine learning models designed to identify acute aortic dissection demonstrate strong overall internal discrimination but experience substantial performance drops depending on comparator composition, according to a retrospective diagnostic study of chest pain patients. The investigated algorithms differentiated acute aortic dissection from acute myocardial infarction with high accuracy, yet their diagnostic performance declined considerably when differentiating acute aortic dissection from acute pulmonary embolism across hospital records.
In clinical settings, acute aortic dissection can closely resemble acute myocardial infarction or acute pulmonary embolism in presenting chest pain patients. Although routine laboratory models have the potential to assist differential diagnosis among these critical conditions, their diagnostic performance may depend substantially on comparator composition. The stated objective of the retrospective investigation was therefore to explore internal discrimination and case-mix dependence within an exploratory, hypothesis-generating study framework.
The retrospective reanalysis examined a total of 977 patient records collected across a two-year period from January 1, 2019, to December 31, 2020. This clinical dataset comprised 276 records from patients diagnosed with acute aortic dissection, 578 records from patients with acute myocardial infarction, and 123 records from patients with acute pulmonary embolism. The study investigators established this final analytical cohort following the formal exclusion of unresolved duplicate or near-duplicate patient records.
To test differential diagnostic performance, the researchers evaluated a D-dimer-only baseline alongside six multivariable machine learning models drawn from a fifty-predictor routine laboratory pool. Model performance was evaluated through a five-fold outer validation procedure repeated ten times, with data preprocessing fitted strictly within the individual training folds. The D-dimer baseline was refitted without parameter tuning, four multivariable models utilized archived fold-specific parameters, and XGBoost and LightGBM underwent new inner hyperparameter tuning. The complete analytical workflow assessed model calibration, decision curves, cardiac-marker ablation, comparator composition, and cross-fitted XGBoost SHAP values.
Across the overall model evaluation, LightGBM recorded an area under the receiver operating characteristic curve of 0.941, with a 95 percent confidence interval of 0.927 to 0.953, and a Brier score of 0.0897. XGBoost achieved an overall area under the curve of 0.940, with a 95 percent confidence interval of 0.926 to 0.952, and a Brier score of 0.0903. The paired difference in area under the curve between LightGBM and XGBoost was 0.0013, with a 95 percent confidence interval between -0.0011 and 0.0035. However, model discrimination varied markedly by comparator type: both algorithms reached an area under the curve of approximately 0.977 against acute myocardial infarction, but dropped to 0.773 for LightGBM and 0.767 for XGBoost against acute pulmonary embolism.
Cardiac-marker ablation demonstrated that removing high-sensitivity cardiac troponin T, creatine kinase-MB, and myoglobin decreased overall area under the curve values to 0.917 for LightGBM and 0.916 for XGBoost. When the analysis equally represented acute myocardial infarction and acute pulmonary embolism controls, overall discrimination declined to 0.875 and 0.872, respectively. The study authors concluded that internal discrimination depended directly on comparator composition, displaying notably weaker diagnostic performance against acute pulmonary embolism. Consequently, these exploratory findings do not establish a clinically deployable screening or exclusion tool, and independent validation in consecutive patients presenting with suspected disease remains required.
Subject of Research: Machine learning diagnostic models for acute aortic dissection
Article Title: Development and internal validation of a machine learning-based diagnostic model for acute aortic dissection in chest pain patients: a retrospective study
Article References: Li, P., Xia, Q., Wu, D., Luo, T., Peng, P., & Liu, Y. (2026). Development and internal validation of a machine learning-based diagnostic model for acute aortic dissection in chest pain patients: a retrospective study. International Journal of Emergency Medicine. https://doi.org/10.1186/s12245-026-01423-7
Image Credits: AI Generated
DOI: 10.1186/s12245-026-01423-7
Keywords: acute aortic dissection, chest pain, differential diagnosis, machine learning, acute myocardial infarction, acute pulmonary embolism, LightGBM, XGBoost, D-dimer, internal validation, cardiac biomarkers, high-sensitivity troponin T, creatine kinase-MB, myoglobin, Brier score, SHAP analysis, case-mix dependence, comparator composition
News Source: Ophelia Keating. (November 17, 2026). Aortic Dissection Diagnostic Models Show Weaker Performance Against Pulmonary Embolism. Scienmag.



