摘要目的：初步制定危重患者急性胃肠损伤分级数字化预测模型。方法采用双通路胃肠音监测系统对2015年4月至2015年11月期间北京解放军总医院重症医学科连续收治的60例危重患者（排除既往有慢性消化道疾病史或胃肠道手术者）的胃肠音进行采集和分析，同时采用急性胃肠损伤（AGI）分级标准进行评估，胃肠音与AGI分级之间的相关性采用Spearman秩相关分析，之后将显著相关的数据标准化处理后进行主成分分析，选取排在前3的主成分进行归一化后通过反向传播神经网络进行BP神经网络训练，构建神经网络模型，从而建立危重患者急性胃肠损伤分级的初步预测模型。结果共采集到60例患者1132次胃肠音和333次AGI分级的数据。其中胃部测得的胃肠音波数量（P ＝0．0005）、胃肠音波时间百分比（P＝0．0004）、胃肠音波最大时间（P＝0．0088）、胃肠音波平均频率（P＝0．0101）及胃肠音波最大频率（P＝0．0025）与AGI分级呈负相关；而肠部测得的胃肠音各指标均AGI分级无显著相关性（均P ＞0.05）。将这5个与AGI分级相关性较强的胃肠音指标进行主成分分析后，筛选出3个主成分进入BP神经网络模型，建立了包含9个隐含中间层的AGI初步神经网络模型，模型拟合度为0.98164。应用该模型预测患者的AGI分级准确率为70.83%。结论根据胃肠音指标构建的AGI分级初步预测模型，可为医护人员进行危重患者AGI分级的初步预测判断提供参考。

abstractsObjective To develop the prediction model of acute gastrointestinal injury (AGI) classification of critically ill patients. Methods The binary channel gastrointestinal sounds (GIS) monitor system was used to gather and analyze the GIS of 60 consecutive critically ill patients who were admitted in Critical Care Medicine of PLA General Hospital from April 2015 to November 2015 (patients with chronic gastrointestinal disease or history of gastrointestinal surgery were excluded ). Meanwhile, the AGI grades were evaluated according to the ESICM guidelines of AGI grading system. Correlations between GIS and AGI classification were examined with Spearman rank correlation. Then principal component analysis was performed on the significantly correlated parameters after standardization. The top 3 post-normalized main components were selected for back-propagation (BP) neural network training to establish primary AGI grade model of critically ill patients based on the neural network model. Results A total of 1 132 GIS and 333 AGI were collected from 60 patients. The number (P = 0.0005), percentage of time (P = 0.0004), mean power (P = 0.0088), maximum power (P = 0.0101) and maximum time (P = 0.0025) of GIS wave from the channel located at the stomach were negatively correlated with the AGI grades , while the parameters of GIS wave from the channel located at the intestine had no significant correlation with the AGI grades (all P > 0.05). Three main components were selected after principal component analysis of these five correlated parameters. An AGI grade network model including 9 hide layers, with a fitting degree of 0.981 64 was built by BP artificial neural network based on the analysis of these three main components of GIS. The accuracy rate of the model to predict the AGI grade was 70.83%. Conclusion The preliminary model based on GIS in classifying AGI grade is established successfully , which can help predict the classification of AGI grade of critically ill patients.