Machine learning for the prediction of preoxygenation technique in trauma

Background Preoxygenation can be achieved best by non-invasive ventilation techniques (NIV).
Objective With the help of machine learning, the decision-making process against or in favour of NIV for preoxygenation in severely injured preclinical patients shall be evaluated.
Methods A registry-based, retrospective analysis in preclinical adult trauma patients in south-western Germany between 2018 to 2020 was conducted. Attributes considered were the initial vital signs, Glasgow Coma Scale, airway devices, administered medication, description of difficult airway, emergency interventions, shock index, age and pre emergency status. A decision tree model (REPTree) and two Bayesian network (BN) were created, one with all and the other with the attributes occurring in the decision tree.
Results 992 datasets with 333 cases of NIV (33%) were identified. Main splitting points in the decision tree model were the attributes rhonchus and bronchial spasm, videolaryngoscopy, respiratory rate, heart rate, age, oxygen saturation and head injury. The area under the receiver operating characteristics was between 0.97 (original BN; 95% CI, 0.96-0.97) and 0.93 (REPTree, 95% CI, 0.92-0.93). For the prediction, the precision-recall area was 0.96 (BN, 95% CI, 0.96-0.97) and 0.88 (REPTree, 95% CI, 0.87-0.89) and for exclusion 0.96 (BN, 95% CI, 0.96-0.97) and 0.94 (REPTree, 65% CI, 0.93-0.94). The simplified BN performed equally to the original BN.
Conclusion The presented models demonstrated a feasibility for modeling decision making as well as an excellent performance. An expended model should contain internal and neurological patients as well as the effectiveness of the chosen method and could therefore support emergency medical crews.

Files:
Supplement Bayesian Network max 3 nodes XML BIF.xml
• XML data file with all nodes and probabilities of the final Bayesian network with a maximum of 3 parental nodes that can be implemented in WEKA
Supplement Simplified Bayesian Network max 3 nodes XML BIF.xml
• XML data file with all nodes and probabilities of the simplified Bayesian network with a maximum of 3 parental nodes that can be implemented in WEKA

背景 预氧合的最优实现方式为非侵入性通气技术（non-invasive ventilation, NIV）。
研究目的 借助机器学习技术，对创伤重症院前患者实施预氧合时，选择采用或放弃非侵入性通气技术的决策流程开展评估。
研究方法 本研究针对2018年至2020年德国西南部的院前成年创伤患者开展基于登记队列的回顾性分析。纳入分析的属性包括初始生命体征、格拉斯哥昏迷量表（Glasgow Coma Scale）、气道装置、给药方案、困难气道描述、急诊干预措施、休克指数、年龄及院前急救状态。本研究构建了决策树模型（REPTree）以及两个贝叶斯网络（Bayesian network, BN），其中一个使用全部属性，另一个仅使用决策树中出现的属性。
研究结果 共纳入992份数据集，其中333例采用非侵入性通气技术（占比33%）。决策树模型的主要分割节点属性包括鼾音、支气管痉挛、视频喉镜检查、呼吸频率、心率、年龄、血氧饱和度及颅脑损伤。受试者工作特征曲线下面积介于0.97（原始贝叶斯网络；95%置信区间CI：0.96~0.97）与0.93（REPTree模型；95%CI：0.92~0.93）之间。针对入选病例的预测，精确召回曲线下面积分别为0.96（贝叶斯网络；95%CI：0.96~0.97）与0.88（REPTree模型；95%CI：0.87~0.89）；针对排除病例的预测则为0.96（贝叶斯网络；95%CI：0.96~0.97）与0.94（REPTree模型；95%CI：0.93~0.94）。简化版贝叶斯网络的性能与原始贝叶斯网络相当。
研究结论 本研究提出的模型证实了决策建模的可行性，且具备优异的预测性能。后续扩展后的模型应纳入内科及神经内科患者群体，同时纳入所选干预方式的有效性指标，从而可为急诊医疗团队提供决策支持。
附件文件：
• 《Supplement Bayesian Network max 3 nodes XML BIF.xml》：可在WEKA软件中运行的最终贝叶斯网络的全节点与概率XML数据文件，该网络最多包含3个父节点。
• 《Supplement Simplified Bayesian Network max 3 nodes XML BIF.xml》：可在WEKA软件中运行的简化版贝叶斯网络的全节点与概率XML数据文件，该网络最多包含3个父节点。