Application of interpretable machine learning algorithms to predict diagnostic yield of radial probe endobronchial ultrasound transbronchial lung biopsy with guide sheath in peripheral pulmonary lesions

Abstract Radial probe endobronchial ultrasound transbronchial lung biopsy with guide sheath (RP-EBUS-GS-TBLB) was one of the main diagnostic methods for peripheral pulmonary lesions (PPLs). The aim of this study was to identify potential factors contributing to the diagnostic yield of RP-EBUS-GS-TBLB in PPLs. A total of 189 consecutive patients with PPLs who had undergone RP-EBUS-GS-TBLB between January 2022 and October 2024 in 8th Medical Centre, Chinese PLA General Hospital were enrolled in this retrospective single-center cohort study. Patients were randomly divided into the training set (70%, n = 132) and test set (30%, n = 57). The most influential factors were selected through the application of Boruta algorithm and LASSO regression method, employing a Venn test for further refinement. Subsequently, five machine learning (ML) techniques, namely XGBoost, LightGBM, Random Forest (RF), Adaptive Boosting (Adaboost), and Support Vector Machine (SVM) were employed to construct predictive models. Model validity was assessed using accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and the area under the receiver operating characteristic curve (AUC). Shapley Additive Interpretation (SHAP) analysis facilitated a visual interpretation for individual patient. The integration of Boruta and LASSO methodologies identified lesion morphology in CT, number of biopsies, size, margin, echogenicity and RP-EBUS location as significant predictors. Taking all factors into consideration, we chose the RF model as the optimal model, with accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and AUC (95% CI) values of 0.857(0.799–0.915), 0.652(0.526–0.779), 0.921(0.856–0.986), 0.759(0.566–0.953), 0.896(0.854–0.937), 0.686(0.573–0.799), and 0.872 (0.695-1.000). SHAP analysis provided independent explanations, reaffirming the critical clinical factors associated with the diagnostic yield of RP-EBUS-GS-TBLB in PPLs. This study successfully established a precise predictive model for diagnostic yield of RP-EBUS-GS-TBLB in PPLs using machine learning techniques, offering valuable support to clinicians in making informed clinical decisions.