MACHINE LEARNING MODELS TO PREDICT FAILURE IN ADMINISTRATION OF INTRAPLEURAL TPA/DNASE IN PATIENTS WITH COMPLICATED PARAPNEUMONIC EFFUSIONS/EMPYEMA: A MULTICENTER STUDY

Abstract

TOPIC: Disorders of the Pleura TYPE: Original Investigations PURPOSE: Intrapleural fibrinolytic therapy (IPFT) with tissue plasminogen activator (tPA) and deoxyribonuclease (DNase) has been shown to reduce the need for surgical intervention for complicated pleural effusion/empyema (CPE/empyema). For patients in whom tPA/DNase is likely to fail, however, receipt of this therapy may simply delay the inevitable. The goal of this study was to identify risk factors for failure of combined intrapleural therapy. METHODS: We performed a multicenter, retrospective chart review of patients who received intrapleural tPA/DNase for the treatment of CPE/empyema. Clinical variables included demographic data, radiographic parameters at time of diagnosis, and results from pleural fluid analysis. We compared four different machine learning classifiers (L1-penalized logistic regression, support vector machine, XGBoost and LightGBM) by multiple bootstrap-validated metrics, including F-beta (with recall weighted twice as heavily as precision) and area under the receiver operator characteristic (AUC) curve. Model hyperparameters were tuned using adaptive optimization (tree-parzan estimator), and variable importance was estimated by random permutation to rank the importance of 19 candidate clinical variables with respect to their ability to predict failure of tPA/DNase therapy. Data were randomly split into training and test sets using an 80:20 ratio. Training data were further randomly split by the same ratio for hyperparameter tuning. The whole pipeline was repeated 200 times in order to derive distributions for evaluation metrics. RESULTS: We included 466 participants from five institutions across the United States. All participants in this study received IPFT with tPA/DNase for the management of complicated pleural effusions/empyema. Resolution of CPE/empyema with intrapleural tPA/DNase was achieved in 78% (n=365) of cases. Support vector machine (SVM) performed the best with F-beta of 57%, followed by XGBoost (F-beta 47%), and LGBM (F-beta 46%). Although the average performance of L1-penalized logistic regression was adequate (50% F-beta), it was highly unstable with F-beta ranging from 0 to 100%. Of the 19 candidate predictors of tPA/DNase failure, all models exceptL1 agreed that the single most important predictor was whether abscess or necrotizing pneumonia were present, followed by the presence of pleural thickening and protein pleural. (L1 ranked the presence of pleural thickening as most important.) CONCLUSIONS: Our analysis based on a large, multicenter database demonstrated the presence of pleural thickening and abscess/necrotizing pneumonia helps to triage patients in whom IPFT is likely to fail and surgical management may be indicated. These results warrant further investigation and validation in a prospective study. CLINICAL IMPLICATIONS: Simple radiographic findings with presence of pleural thickening and abscess/necrotizing pneumonia can be used to triage patients in whom intrapleural fibrinolytic therapy is likely to fail. These two variables may be helpful to expedite surgical intervention in patients with pleural infection. DISCLOSURES: No relevant relationships by Akshu Balwan, source=Web Response No relevant relationships by Billie Bixby, source=Web Response Consultant relationship with olympus medical inc Please note: $1-$1000 by Christopher Gilbert, source=Web Response, value=Consulting fee Consultant relationship with verathon Please note: $1001 - $5000 by Christopher Gilbert, source=Web Response, value=Consulting fee Consultant relationship with AstraZenaca Please note: $5001 - $20000 by Christopher Gilbert, source=Web Response, value=Consulting fee no disclosure on file for Jed Gordon; No relevant relationships by Danai Khemasuwan, source=Web Response no disclosure on file for Chakravarthy Reddy; No relevant relationships by Trinidad Sanchez, source=Web Response No relevant relationships by Samira Shojaee, source=Web Response No relevant relationships by Jeff Sorensen, source=Web Response No relevant relationships by Candice Wilshire, source=Web Response