Machine Learning Model Predictors of Intrapleural Tissue Plasminogen Activator and DNase Failure in Pleural...

Rationale: Intrapleural enzyme therapy (IET) with tissue plasminogen activator (tPA) and DNase has been shown to reduce the need for surgical intervention for complicated parapneumonic effusion/empyema (CPPE/empyema). Failure of IET may lead to delayed care and increased length of stay. Objectives: The goal of this study was to identify risk factors for failure of IET. Methods: We performed a multicenter, retrospective study of patients who received IET for the treatment of CPPE/empyema. Clinical and radiological variables at the time of diagnosis were included. We compared four different machine learning classifiers (L1-penalized logistic regression, support vector machine [SVM], extreme gradient boosting [XGBoost], and light gradient-boosting machine [LightGBM]) by multiple bootstrap-validated metrics, including F-β, to demonstrate model performances. Results: A total of 466 participants who received IET for pleural infection were included from five institutions across the United States. Resolution of CPPE/empyema with IET was achieved in 78% (n = 365). SVM performed superiorly, with median F-β of 56%, followed by L1-penalized logistic regression, LightGBM, and XGBoost. Clinical and radiological variables were graded based on their ranked variable importance. The top two significant predictors of IET failure using SVM were the presence of an abscess/necrotizing pneumonia (17%) and pleural thickening (13%). Similarly, LightGBM identified abscess/necrotizing pneumonia (35%) and pleural thickening (26%) and XGBoost indicated pleural thickening (36%) and abscess/necrotizing pneumonia (17%) as the most significant predictors of treatment failure. Predictors identified by the L1-penalized logistic regression model were pleural thickening (18%) and pleural fluid lactate dehydrogenase (LDH) (9%). Conclusions: The presence of abscess/necrotizing pneumonia and pleural thickening consistently ranked among the strongest predictors of IET failure in all machine learning models. The difference in rankings between models may be a consequence of the different algorithms used by each model. These results indicate that the presence of abscess/necrotizing pneumonia and pleural thickening may predict IET failure. These results should be confirmed in larger studies.