Abstract No. 25 ▪ FEATURED ABSTRACT Use of machine learning to predict the risk of pneumothorax requiring chest tube placement after lung biopsy

Abstract

Pneumothorax is the leading complication after image-guided lung nodule biopsy (15-25%) with 4-6% resulting in a pneumothorax requiring chest tube placement . Risk factors associated with lung biopsy induced pneumothorax have been elucidated; however, currently no model exists to prospectively identify patients most at risk for pneumothorax necessitating chest tube placement. Development of a machine learning algorithm to predict patients at high risk for pneumothorax and chest tube placement may facilitate appropriate peri-procedural planning and optimize patient care. Retrospective chart review was performed on patients who underwent lung nodule biopsy between January 2018 and July 2019. From this data, we selected 58 lung biopsy procedures that resulted in pneumothorax requiring chest tube placement and then another 51 procedures without chest tube placement (convenience sample). For each procedure, we collected 12 clinical features that would be available prior to biopsy: age, gender, smoking history, prior asthma or COPD diagnosis, cancer history, pulmonary function tests, biopsy needle gauge, lesion size, lesion location, lesion character (solid or not), patient position during biopsy, and pleura to lesion distance. The data was divided 70/30 into training and testing sets. We explored several conventional machine-learned binary classifiers with the clinical data as input features and need for chest tube placement as the output. Summary statistics of the data reveal that several of the features have unique means between the chest tube placement positive and negative groups. For example, the set requiring tube placement versus not had a mean lesion depth of 30 ± 19 mm versus 13 ± 14 mm (P< 0.005). That said, there was no single pathognomonic imaging or clinical feature. On the independent test set, use of decision tree, logistic regression, and Naïve Bayes achieved accuracies of identifying tube placement of 0.76, 0.74, and 0.76 and had AUC’s of 0.83, 0.76, 0.78. The learned features of high importance included lesion character, COPD status, lesion depth, and patient age. A coarse decision tree requiring only 3 inputs achieved comparable performance as other methods. We identified an interpretable decision tree to predict chest tube placement post biopsy with an accuracy of 76% and AUC of 0.83. This work suggests that it may be possible to predict the risk of pneumothorax requiring chest tube placement based on an automated decision support tool reliant on easily available clinical and imaging features.