Accurate Differentiation of Acute Aortic Syndrome from Acute Coronary Syndrome: A Machine Learning Diagnostic Model Based on Routine Admission Variables

Background/Objectives: Acute aortic syndrome (AAS) is an underdiagnosed, time-critical medical emergency. Initial misdiagnosis as acute coronary syndrome (ACS) is common, delaying definitive treatment and exposing patients to inappropriate antithrombotic therapy. We aimed to develop and temporally test a machine learning (ML) model using routine admission variables to distinguish AAS from ACS. Methods: In this single-center retrospective study, we included 2828 consecutive adults with confirmed AAS or ACS (September 2018–April 2025); the initial 1979 patients formed the development cohort and the subsequent 849, a temporally separated test cohort. Least absolute shrinkage and selection operator regression selected predictors from nine prespecified candidates. Seven algorithms were fitted and assigned fixed Youden-derived cutoffs in the development cohort, then applied to the test cohort. Results: Six predictors were selected: D-dimer, leukocyte count, troponin I, C-reactive protein, oxygen saturation, and ST-segment changes. In the test cohort (112 AAS [13.2%]; 737 ACS), area under the receiver operating characteristic curve (AUC) ranged from 0.925 to 0.956 across the seven models. Extreme gradient boosting (XGBoost) showed the highest observed AUC (0.956; 95% confidence interval (CI), 0.935–0.976) and the lowest Brier score (0.047; 95% CI, 0.037–0.057), with sensitivity 0.884, specificity 0.931, and negative predictive value 0.981 at the development-cohort-derived cutoff (0.199); decision curve analysis indicated net benefit within the AAS/ACS test cohort. Shapley additive explanations ranked D-dimer as the dominant contributor. Conclusions: Six routine admission variables discriminated AAS from ACS across all seven ML algorithms in this single-center cohort. These findings support further prospective evaluation in a broader diagnostic population.