Development of an interpretable machine learning model for early prediction of aortic stiffness risk in health examination populations

Background: Carotid-femoral pulse wave velocity (cfPWV) is the gold standard for assessing aortic stiffness, but its complexity, time consumption, and privacy concerns limit its application in routine health examinations. This study aimed to develop an interpretable machine learning model based on readily accessible indicators, providing an alternative screening tool for institutions unable to perform cfPWV measurements. Methods: A total of 261 participants were enrolled as the development cohort and randomly divided into training and testing sets (7:3 ratio), with 101 additional participants as an independent external validation set. Aortic stiffness was defined as cfPWV ≥10 m/s. Features were selected by combining univariate logistic regression, recursive feature elimination with random forest (RFE-RF), and least absolute shrinkage and selection operator (LASSO) methods. Six machine learning models were constructed: logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), Naive Bayes (NB), and K-nearest neighbours (KNN). Internal validation was performed using 5-fold cross-validation, and performance was evaluated using area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). The optimal model was validated on the external set, and interpretability was analyzed using SHapley Additive exPlanations (SHAP). Subsequently, the model was deployed as an interactive, web-based application utilizing the Streamlit framework in Python. Results: Seven key variables were selected: Age, body mass index (BMI), mean arterial pressure (MAP), fasting blood glucose (FBG), high-density lipoprotein cholesterol (HDL-C), glomerular filtration rate (GFR), and aspartate aminotransferase (AST). The XGBoost model achieved the best performance with an AUC of 0.903 (95% CI: 0.830-0.975) on the testing set and a mean AUC of 0.979 (95% CI: 0.960-0.997) in 5-fold cross-validation on the training set. External validation demonstrated robust generalizability (AUC = 1.000). DCA indicated a favourable clinical net benefit, and SHAP analysis quantified feature contributions. Conclusion: This study developed a high-performance XGBoost model based on routine health examination indicators and further implemented an accessible web-based calculator to support clinical decision-making in settings without direct access to cfPWV measurement. The calculator is available at: https://lgezijo5wyivqrnt2zrkcm.streamlit.app/.