Development and internal validation of a machine learning-based diagnostic model for acute aortic dissection in chest pain patients: a retrospective study

Acute aortic dissection (AAD) can resemble acute myocardial infarction (AMI) or acute pulmonary embolism (APE). Routine laboratory models may assist differential diagnosis, but their performance may depend on comparator composition. Objective: To explore internal discrimination and case-mix dependence in a hypothesis-generating study. The reanalysis included 977 records from January 1, 2019 to December 31, 2020 (276 AAD, 578 AMI and 123 APE), after exclusion of unresolved duplicate or near-duplicate records. A D-dimer-only baseline and six multivariable models using a 50-predictor laboratory pool were evaluated by five-fold outer validation repeated ten times. Preprocessing was fitted within training folds. The D-dimer baseline was refitted without tuning; four models used archived fold-specific parameters; XGBoost and LightGBM underwent new inner tuning. Analyses assessed calibration, decision curves, cardiac-marker ablation, comparator composition and cross-fitted XGBoost SHAP. Overall AUC was 0.941 (95% CI 0.927–0.953) for LightGBM and 0.940 (0.926–0.952) for XGBoost; Brier scores were 0.0897 and 0.0903. Their paired AUC difference was 0.0013 (− 0.0011 to 0.0035). AUCs were approximately 0.977 against AMI but 0.773 and 0.767 against APE. Removing hs-cTnT, CK-MB and myoglobin reduced overall AUC to 0.917 and 0.916; equally representing AMI and APE controls reduced it to 0.875 and 0.872. Internal discrimination depended on comparator composition, with weaker performance against APE. These exploratory findings do not establish a clinically deployable screening or exclusion tool. Independent validation in consecutive patients with suspected disease is required.