A machine learning model for predicting pneumothorax risk after computed tomography-guided percutaneous transthoracic...

Pneumothorax is a common complication after computed tomography-guided percutaneous transthoracic needle biopsy (PTNB), but individualized prediction tools based on routinely available clinical variables remain limited. This two-hospital retrospective study included 768 patients who underwent CT-guided PTNB between January 2024 and August 2025. The pooled dataset was randomly divided into a training cohort and an internal test cohort at a 7:3 ratio. Twelve predictors were selected using LASSO regression. Seven machine-learning models were developed and evaluated using discrimination, calibration, decision curve analysis, and SHAP-based interpretability. The support vector machine (SVM) model showed relatively balanced internal performance compared with the other candidate models. In the internal test cohort, the SVM model achieved an AUC of 0.798, accuracy of 0.766, sensitivity of 0.881, specificity of 0.400, positive predictive value of 0.824, negative predictive value of 0.512, F1-score of 0.852, and PR-AUC of 0.929. Calibration analysis showed acceptable agreement between predicted probabilities and observed pneumothorax incidence. Decision curve analysis suggested potential clinical net benefit across the evaluated threshold range. SHAP analysis identified smoking history, age, and lactate dehydrogenase as important contributors to model prediction.The final selected SVM model showed relatively balanced internal performance for predicting pneumothorax risk after CT-guided PTNB. This model may support pre-procedural risk assessment and individualized post-procedural monitoring, but center-based and prospective external validation is needed.