Dente Christopher J, Bradley Matthew, Schobel Seth, Gaucher Beverly, Buchman Timothy, Kirk Allan D, Elster Eric
From the Emory University (C.J.D., T.B.), Atlanta, Georgia; Grady Memorial Hospital (C.J.D.), Atlanta, Georgia; Uniformed Services University of the Health Sciences (M.B., S.S., B.G., E.E.), Bethesda, Maryland; Walter Reed National Military Medical Center (M.B., E.E.), Bethesda, Maryland; Surgical Critical Care Initiative (SC2i) (C.J.D., M.B., S.S., B.G., T.B., A.D.K., E.E.), Bethesda, Maryland; and Duke University (A.D.K.), Durham, North Carolina.
J Trauma Acute Care Surg. 2017 Oct;83(4):609-616. doi: 10.1097/TA.0000000000001596.
The biomarker profile of trauma patients may allow for the creation of models to assist bedside decision making and prediction of complications. We sought to determine the utility of modeling in the prediction of bacteremia and pneumonia in combat casualties.
This is a prospective, observational trial of patients with complex wounds treated at Walter Reed National Military Medical Center (2007-2012). Tissue, serum, and wound effluent samples were collected during operative interventions until wound closure. Clinical, biomarker, and outcome data were used in machine learning algorithms to develop models predicting bacteremia or pneumonia. Modeling was performed on the first operative washout to maximize predictive benefit. Variable selection of dataset variables was performed and the best-fitting Bayesian belief network (BBN), using Bayesian information criterion (BIC), was selected for predictive modeling. Random forest was performed using variables from BBN step. Model performance was evaluated using area under the receiver operating characteristic curve (AUC) analysis.
Seventy-three patients (mean age 23, mean Injury Severity Score 25) were enrolled. Patients required a median of 3 (2-13) operations. The incidence of bacteremia and pneumonia was 22% and 12%, respectively. Best-fitting variable selected BBNs were maximum-minimum parents and children (MMPC) for both bacteremia (BIC-24948) and pneumonia (BIC-17886). Full variable and MMPC random forest models AUC were 0.721 and 0.834, respectively, for bacteremia and 0.809 and 0.856, respectively, for pneumonia.
We identified a profile predictive of bacteremia and pneumonia in combat casualties. This has important clinical implications and should be validated in the civilian trauma population. This and similar tools will allow for increasing precision in the management of critically ill and injured patients.
Prognostic, level III.
创伤患者的生物标志物特征可能有助于建立模型，以辅助床边决策和预测并发症。我们试图确定建模在预测战斗伤员菌血症和肺炎方面的效用。
这是一项对在沃尔特里德国家军事医疗中心接受治疗的复杂伤口患者进行的前瞻性观察性试验（2007 - 2012年）。在手术干预期间直至伤口闭合，收集组织、血清和伤口流出液样本。临床、生物标志物和结局数据用于机器学习算法，以建立预测菌血症或肺炎的模型。在首次手术冲洗时进行建模，以最大化预测效益。对数据集变量进行变量选择，并使用贝叶斯信息准则（BIC）选择最佳拟合的贝叶斯信念网络（BBN）进行预测建模。使用来自BBN步骤的变量进行随机森林分析。使用受试者操作特征曲线（AUC）分析评估模型性能。
共纳入73例患者（平均年龄23岁，平均损伤严重程度评分25分）。患者平均需要3次（2 - 13次）手术。菌血症和肺炎的发生率分别为22%和12%。对于菌血症（BIC - 24948）和肺炎（BIC - 17886），选择的最佳拟合变量BBN分别是最大 - 最小父节点和子节点（MMPC）。对于菌血症，完整变量和MMPC随机森林模型的AUC分别为0.721和0.834，对于肺炎分别为0.809和0.856。
我们确定了一种可预测战斗伤员菌血症和肺炎的特征。这具有重要的临床意义，应在 civilian创伤人群中进行验证。这一工具及类似工具将提高对重症和受伤患者管理的精准度。
预后性，III级。
