Background
Tools to assist clinicians in predicting pneumonia could lead to a significant decline in morbidity. Therefore, we sought to develop a model in combat trauma patients for identifying those at highest risk of pneumonia.
Methods
This was a retrospective study of 73 primarily blast‐injured casualties with combat extremity wounds. Binary classification models for pneumonia prediction were developed with measurements of injury severity from the Abbreviated Injury Scale (AIS), transfusion blood products received before arrival at Walter Reed National Military Medical Center (WRNMMC), and serum protein levels. Predictive models were generated with leave‐one‐out‐cross‐validation using the variable selection method of backward elimination (BE) and the machine learning algorithms of random forests (RF) and logistic regression (LR). BE was attempted with two predictor sets: (1) all variables and (2) serum proteins alone.
Results
Incidence of pneumonia was 12% (n = 9). Different variable sets were produced by BE when considering all variables and just serum proteins alone. BE selected the variables ISS, AIS chest, and cryoprecipitate within the first 24 h following injury for the first predictor set 1 and FGF‐basic, IL‐2R, and IL‐6 for predictor set 2. Using both variable sets, a RF was generated with AUCs of 0.95 and 0.87—both higher than LR algorithms.
Conclusion
Advanced modeling allowed for the identification of clinical and biomarker data predictive of pneumonia in a cohort of predominantly blast‐injured combat trauma patients. The generalizability of the models developed here will require an external validation dataset.
IntroductionVenous thromboembolism (VTE) is a frequent complication of trauma associated with high mortality and morbidity. Clinicians lack appropriate tools for stratifying trauma patients for VTE, thus have yet to be able to predict when to intervene. We aimed to compare random forest (RF) and logistic regression (LR) predictive modelling for VTE using (1) clinical measures alone, (2) serum biomarkers alone and (3) clinical measures plus serum biomarkers.MethodsData were collected from 73 military casualties with at least one extremity wound and prospectively enrolled in an observational study between 2007 and 2012. Clinical and serum cytokine data were collected. Modelling was performed with RF and LR based on the presence or absence of deep vein thrombosis (DVT) and/or pulmonary embolism (PE). For comparison, LR was also performed on the final variables from the RF model. Sensitivity/specificity and area under the curve (AUC) were reported.ResultsOf the 73 patients (median Injury Severity Score=16), nine (12.3%) developed VTE, four (5.5%) with DVT, four (5.5%) with PE, and one (1.4%) with both DVT and PE. In all sets of predictive models, RF outperformed LR. The best RF model generated with clinical and serum biomarkers included five variables (interleukin-15, monokine induced by gamma, vascular endothelial growth factor, total blood products at resuscitation and presence of soft tissue injury) and had an AUC of 0.946, sensitivity of 0.992 and specificity of 0.838.ConclusionsVTE may be predicted by clinical and molecular biomarkers in trauma patients. This will allow the development of clinical decision support tools which can help inform the management of high-risk patients for VTE.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.