D.J. Puccio1, J. Hörauf1, A. Rosen1, U. Kar1, R. Zamora1, J. Sperry1, J. Das1, T. Billiar1 1University Of Pittsburgh, Department Of Surgery, Pittsburgh, PA, USA
Introduction: Trauma outcome predictive models have historically included demographics, injury severity scores, and standard-of-care laboratory values. Identifying key predictive biomarkers is challenging due to correlations among acute blood chemokine and cytokines. Statistical models that have a high level of prediction for adverse outcomes measured early in the clinical course (especially admission) have been lacking in part due to limitations in feature selection methods. Stabl is a recently described method for feature selection using knockoffs to control for false discovery rate, without definition of an a priori selection threshold, and allows us to identify a small group of robust and reliable predictive biomarkers. The purpose of this study was to apply Stabl to a dataset of 21 immune mediators measured at admission (0hr) and 24hrs to predict adverse outcomes in a large clinical dataset of severely injured polytrauma patients.
Methods: We examined a prospectively enrolled dataset from the Prehospital Air Medical Plasma (PAMPer) trial which included clinical data and blood samples collected at 0hr and 24hrs following emergency room admission. Samples were analyzed by a blinded technician using a 21 chemokine/cytokine Luminex assay. Patient outcomes were segregated into those that died in the first first 72hrs or had prolonged intensive care unit (ICU) length of stay (LOS) vs rapid recovery (ICU LOS < 7 days). From our 0-hour chemokine/cytokine data, Stabl identifies a small group of chemokines and cytokines to predict outcome, which we then validate using k-fold cross validation. Using this fitted model, we evaluate the performance of these chemokines and cytokines using data at 24hrs.
Results: The retrospective dataset contained 501 trauma patients, with 375 patients having completed outcome and chemokine/cytokine data. Our 0hr model selected 5 predictive features for clinical course: monocyte chemoattractant protein-1 (MCP-1), interleukin 10 (IL-10), granulocyte-macrophage colony-stimulating factor (GM-CSF), interleukin 7 (IL-7), and interleukin 21 (IL-21). The receiver operator curve (ROC) analysis yielded area under the curve (AUC) of 0.823 (95% confidence interval (CI): 0.766 – 0.873) for the training set and 0.760 (CI: 0.649 – 0.863) for the testing set. MCP-1, IL-10, and IL-7 remained predictive at 24hrs (Training AUC: 0.816 (CI: 0.762 – 0.863), Testing AUC: 0.748 (CI: 0.690 – 0.803)).
Conclusion: The use of Stabl as a statistical method in this polytrauma dataset allowed feature selection a sparse set of immune mediators, with stringent control for false discovery rate. The chemokine and cytokines MCP-1, IL-10, GM-CSF, IL-7, IL-21 are key biomarkers associated with trauma outcome. Our model highlights sparse immune mediators from a larger biomarker pool to predict adverse outcomes; similar emerging computational tools may assist clinicians in prognostication using noisy and highly correlated clinical features.