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
Retrospective cohort study demonstrates that machine learning reliably predicts bronchoscopy biopsy yield in peripheral pulmonary lesions, indicating potential for personalized clinical...
Key Points
To identify key determinants of diagnostic yield and develop interpretable machine learning models predicting biopsy success in peripheral pulmonary lesions evaluated with radial probe endobronchial ultrasound.
Conducted a retrospective single-center cohort study evaluating 189 consecutive patients with peripheral pulmonary lesions randomly partitioned into a training set (70%, n = 132) and a test set (30%, n = 57).
Selected predictive features using the Boruta algorithm and LASSO regression, then developed five machine learning models (XGBoost, LightGBM, Random Forest, AdaBoost, and Support Vector Machine) evaluated using SHAP analysis.
Boruta and LASSO feature selection determined that CT lesion morphology, biopsy count, lesion size, margin, echogenicity, and ultrasound probe location were the primary predictors of diagnostic yield.
The Random Forest model demonstrated the best overall performance, achieving an AUC of 0.872 (95% CI: 0.695–1.000), accuracy of 0.857 (95% CI: 0.799–0.915), sensitivity of 0.652 (95% CI: 0.526–0.779), and specificity of 0.921 (95% CI: 0.856–0.986).