Interpretable Machine Learning Approaches for Identification of Acute Aortic Dissection in Chest Pain Patients

The aim of this study is using interpretable machine learning (ML) methods to construct models by combing routine laboratory examination biomarkers and clinical characteristics to identify acute aortic dissection (AAD) patients from other sudden chest pain patients referring to acute myocardial infarction (AMI), acute pulmonary embolism, and abdominal aortic aneurysm.

The research encompassed a cohort of 832 individuals, with 515 of them diagnosed as AAD patients. Patients were randomly assigned to training and test groups for model development and evaluation, with data collected from medical records and validated by study physicians. Logistic Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for variable selection in the study, which utilized 9 ML algorithms for model development. The DeLong test compared area under the curve (AUC) values among models. Optimal parameters were found through grid search on the training set with 5-fold cross validation. The SHapley Additive exPlanation (SHAP) method ranks input feature importance and explains model outcomes to address model opacity.

Utilizing the LASSO regression technique, 8 variables were pinpointed for their nonlinear significance. Evaluation of these models using test set data yielded AUC values between 0.72 and 0.77, suggesting promising utility in differential diagnosis. The Random Forest method demonstrated noteworthy sensitivity, specificity, and F1 Score. The internal validation set consistently yielded results with an AUC ranging from 0.71 to 0.77. The SHAP method was utilized to assess the influence of features on the model, identifying neutrophil (N.L) and age as the most significant variables.

In this prognostic study, a ML model was created to assist in differentiating patients with aortic dissection from those presenting with chest pain. The use of interpretable ML techniques allows for the prioritization of key features, showcasing significant potential for application in supporting the prompt diagnosis and treatment of aortic dissection differentials.