A NEURAL NETWORK APPROACH TO PEDIATRIC TRAUMA TRIAGE

描述（由申请人提供）：伤害导致的儿童死亡人数超过所有其他原因的总和。由于受伤儿童在有专门创伤护理的中心有更好的预后，因此适当的运送到这些中心可能会降低发病率和死亡率。院前信息必须用于确定是否需要运送到创伤中心并规划到达后的护理。根据院前信息评估了几种分诊方案，以对轻度和重度受伤儿童进行分层。这些研究使用了各种各样的输入，包括生理参数、损伤解剖部位和损伤机制，并使用了线性分析和逻辑回归技术。尽管做出了这些努力，但没有一种儿科创伤分诊方法达到足够准确（避免分诊不足或过度）和可重复性的目标。使用标准回归和分类技术制定的分诊标准有限成功的一个潜在解释是院前数据的复杂性以及这些数据与结果的不确定关系。当分类需要使用不同类型的输入数据，或者这些变量与最终分类之间的关系模糊不清时，神经网络是一组比传统方法更有优势的模型。虽然它们在分诊中的应用尚未被描述，但神经网络在预测成人创伤患者的发病率和死亡率方面优于逻辑回归方法。由于院前变量的多样性和年龄特异性，以及这些变量与受伤儿童结局之间的不确定关系，神经网络可能非常适合于分析儿科创伤分诊。这些观察结果导致了一个假设，即神经网络可以比目前的分诊方法更准确地根据损伤的严重程度和对医院资源的需求对儿科创伤患者进行分类。神经网络将接受训练，根据受伤的严重程度和对医院资源的需求，利用院前提供者可获得的信息，对受伤儿童进行分类。将训练后网络的分类精度与基于当前分类策略的分类精度进行比较。该项目的长期目标是开发和实施一种基于神经网络的受伤儿童分诊工具。

DESCRIPTION (provided by applicant): Injuries result in more deaths in children than all other causes combined. Because injured children have better outcomes at centers with specialized trauma care, appropriate transport to these centers may reduce morbidity and mortality. Prehospital information must be used to determine the need for transport to a trauma center and for planning care after arrival. Several triage schemes have been evaluated for stratifying minimally and severely injured children based on prehospital information. These have used a diverse range of inputs, including physiological parameters, anatomic site of injury and mechanism of injury, and have used both linear techniques of analysis and logistic regression. Despite these efforts, no pediatric trauma triage method has yet met the goal of being sufficiently accurate (avoiding under- or over-triage) and reproducible. One potential explanation for the limited success of triage criteria developed using standard regression and classification techniques is the complexity of prehospital data and the uncertain relationship of this data to outcome. Neural networks are a family of models that have an advantage over conventional methods when classification requires using different types of input data or the relationships between these variables and the final classification are vaguely understood. While their use for triage has not been described, neural networks have been better than logistic regression methods in predicting morbidity and mortality in adult trauma patients. Neural networks may be well suited for analysis of pediatric trauma triage because of the diversity and age-specificity of prehospital variables and the uncertain relationship between these variables and the outcome of injured children. These observations have led to the hypothesis that a neural network can more accurately classify pediatric trauma patients by severity of injury and need for hospital resources than current triage methods. Neural networks will be trained to classify injured children based on severity of injury and need for hospital resources using information available to prehospital providers. The classification accuracy of trained networks will be compared with that based on current triage strategies. The long-term goal of this project is to develop and implement a tool for triaging injured children based on neural networks.