An explainable machine learning model for type B aortic dissection identification: development and internal...

Type B aortic dissection (TBAD) is a life-threatening cardiovascular emergency that requires timely recognition. This study aimed to develop and internally evaluate an explainable machine learning model for identifying existing TBAD using routinely available clinical history and admission laboratory data.

This single-center retrospective case-control study included 1,640 participants, comprising 854 patients with CTA-confirmed TBAD and 786 hospitalized controls who underwent whole-aorta CTA and were confirmed not to have aortic dissection. Demographic characteristics, medical history, and admission laboratory test results were collected as 38 initial candidate features. Least absolute shrinkage and selection operator (LASSO) regression was applied to identify key discriminative variables. Machine-learning models based on eight algorithms were then developed and compared: support vector machine (SVM), gradient boosting machine (GBM), neural network, extreme gradient boosting (XGBoost), k-nearest neighbors (KNN), adaptive boosting (AdaBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost). Hyperparameters were tuned in the training set using grid search and repeated 10-fold cross-validation with five repeats. Model discrimination was evaluated in the held-out internal test set using the area under the receiver operating characteristic curve (AUC). The final representative model was selected for exploratory interpretation using SHapley Additive exPlanations (SHAP).

LASSO regression identified five predictors: hypertension, white blood cell count, lymphocyte percentage, basophil percentage, and monocyte count. Several models showed comparable discrimination in the held-out internal test set. The neural network model achieved an AUC of 0.883 in the training set and 0.852 in the held-out internal test set, with 95% confidence intervals of 0.864-0.902 and 0.819-0.885, respectively. Based on the ROC curve analysis, the neural network model showed relatively favorable discrimination and was therefore selected as the final representative model for exploratory SHAP-based interpretation. SHAP analysis indicated that lymphocyte percentage, hypertension, monocyte count, white blood cell count, and basophil percentage were the major contributors to the predictions of the final model, with lymphocyte percentage showing the highest mean absolute SHAP value.

This study developed and internally evaluated an explainable machine-learning model for identifying existing TBAD in a single-center CTA-confirmed retrospective case-control cohort. Given the lack of external validation, this study should be regarded as exploratory. External validation in clinically relevant acute symptomatic populations is required before clinical implementation.