Sullivan Travis M, Kim Mary S, Sippel Genevieve J, Gestrich-Thompson Waverly V, Melhado Caroline G, Griffin Kristine L, Moody Suzanne M, Thakkar Rajan K, Kotagal Meera, Jensen Aaron R, Burd Randall S
Division of Trauma and Burn Surgery, Children's National Hospital, Washington, DC, USA.
Department of Surgery, University of California San Francisco, San Francisco, CA, USA.
J Pediatr Surg. 2025 Feb;60(2):161888. doi: 10.1016/j.jpedsurg.2024.161888. Epub 2024 Aug 31.
Inadequate airway management can contribute to preventable trauma deaths. Current machine learning tools for predicting intubation in trauma are limited to adult populations and include predictors not readily available at the time of patient arrival. We developed a Bayesian network to predict intubation in injured children and adolescents using observable data available upon or immediately after patient arrival.
We obtained patient demographic, injury, resuscitation, and transportation characteristics from trauma registries from four American College of Surgeons-verified level 1 pediatric trauma centers from January 2010 through December 2021. We trained and validated a Bayesian network to predict emergent intubation after pediatric injury. We evaluated model performance using the area under the receiver operating and calibration curves.
The final model, TITAN (Timing of Intubation in Trauma Analysis Network), incorporated five factors: Glasgow Coma Scale, mechanism of injury, injury type (e.g., penetrating, blunt), systolic blood pressure, and age. The model achieved an area under the receiver operating characteristic curve of 0.83 (95% CI 0.80, 0.85) and had a calibration curve slope of 0.98 (95% CI 0.67, 1.29). TITAN had high specificity (98%), negative predictive value (97%), and accuracy (96%) at a binary probability threshold of 22.6%.
The TITAN Bayesian network predicts the risk of intubation in pediatric trauma patients using five factors that are observable early in trauma resuscitation. Prospective validation of the model performance with patient outcomes is needed to assess real-life application benefits and risks.
Prognostic and Epidemiological, Level III.
气道管理不当可导致可预防的创伤死亡。目前用于预测创伤患者插管的机器学习工具仅限于成人，且包括患者到达时无法轻易获得的预测指标。我们开发了一种贝叶斯网络，使用患者到达时或到达后立即可用的可观察数据来预测受伤儿童和青少年的插管情况。
我们从四个经美国外科医师学会验证的一级儿科创伤中心的创伤登记处获取了2010年1月至2021年12月期间患者的人口统计学、损伤、复苏和转运特征。我们训练并验证了一个贝叶斯网络，以预测儿科损伤后的紧急插管情况。我们使用受试者操作曲线下面积和校准曲线评估模型性能。
最终模型TITAN（创伤分析网络中的插管时机）纳入了五个因素：格拉斯哥昏迷量表、损伤机制、损伤类型（如穿透伤、钝器伤）、收缩压和年龄。该模型的受试者操作特征曲线下面积为0.83（95%CI 0.80，0.85），校准曲线斜率为0.98（95%CI 0.67，1.29）。在二元概率阈值为22.6%时，TITAN具有高特异性（98%）、阴性预测值（97%）和准确性（96%）。
TITAN贝叶斯网络使用创伤复苏早期可观察到的五个因素来预测儿科创伤患者的插管风险。需要通过患者结局对模型性能进行前瞻性验证，以评估实际应用的益处和风险。
预后和流行病学，三级。
