3:54 PM Abstract No. 152 Factors associated with pneumothorax in 3719 computed tomography–guided lung biopsies: a machine learning pipeline approach

Abstract

To identify patient and procedural characteristics associated with incidence of pneumothorax in CT-guided lung biopsies using a machine learning pipeline. A total of 3719 CT-guided percutaneous lung biopsies performed at a single institution on 3493 patients (age: 18 to 100 years; 55% male and 45% female) between January 1, 2008, and December 31, 2018, were identified using quality initiative procedure database search after obtaining IRB exemption. 14 patient and procedural characteristics (features) were analyzed. Data were prefitted to a Support Vector Machine filter and recursive feature elimination (SVM-RFE) was used to exclude 5 features. The resulting data with 9 features were analyzed by an ensemble of three machine learning algorithms: Ridge regression, Random Forest classifier, and Gradient Boosting classifier. Seven features were selected for further analysis based on consensus from this ensemble. Chi-squared test was used to determine statistical significance of categorical features (3/7), and discrete simple logistic regression was used for ordinal and continuous features (4/7). Overall incidence of pneumothorax was 6.1%. Age, mean arterial pressure, needle size, number of passes, gender, steroid usage, and involvement of trainee were identified as significant features via recursive feature elimination (SVM-RFE) followed by machine learning ensemble vote. Within categorical features, involvement of trainee was statistically significant (72/228 with trainee vs. 156/228 without trainee; P = 0.026). Within ordinal and continuous features, needle size (P = 0.003) and age (P = 0.03) were statistically significant. Use of steroids, gender, mean arterial pressure, and number of passes were not significantly associated with incidence of pneumothorax. Overall incidence of pneumothorax post CT-guided lung biopsy was lower than previously reported. Machine learning pipelines can help filter large complex datasets to identify characteristics associated with pneumothorax. While this does not imply causation, the findings highlight areas in need of future research, particularly in cases of pneumothorax requiring hospitalization or chest tube placement and complex pulmonary conditions.