An explainable machine learning model based on quantitative HRCT parameters for identifying bronchoscopically confirmed airway stenosis in tracheobronchial tuberculosis

Abstract

Airway stenosis is a frequent and serious complication of tracheobronchial tuberculosis (TBTB), leading to chronic airflow obstruction and reduced quality of life. However, no individualized predictive model currently exists to assess airway narrowing risk in these patients. This retrospective study included 476 patients with TBTB admitted to Suining Central Hospital between January 2017 and July 2024. Demographic, clinical, and quantitative imaging features were analyzed using seven machine learning algorithms. Model training employed repeated cross-validation with grid search for hyperparameter optimization. Model performance was evaluated by discrimination, calibration, and clinical utility, and Shapley Additive Explanations (SHAP) were used for interpretation. The XGBoost model achieved the best overall performance, with an AUC of 0.935, accuracy of 0.930, sensitivity of 0.805, specificity of 0.980. SHAP analysis revealed that the mean T/D ratio and the T/D ratios of the right superior lobar bronchus, right middle lobar bronchus, and left superior lobar bronchus were the most influential imaging predictors, while female sex and symptom duration more than 4 weeks were significant clinical risk factors. The XGBoost model demonstrated superior accuracy, calibration, and interpretability, providing a promising tool for individualized risk assessment airway stenosis in TBTB management.