危重患者急性胃肠损伤分级综合预测模型的构建

Establishment of comprehensive prediction model of acute gastrointestinal injury classification of critically ill patients

目的制定综合危重患者急性胃肠损伤（AGI）分级数字化预测模型。方法2015年4—11月期间，对解放军总医院重症医学科连续收治的60例危重患者以课题组前期研制并验证过准确性的双通路胃肠音监测系统采集胃肠音，进行AGI分级评估（Ⅰ～Ⅳ级，级别越高，胃肠功能障碍越重）；同时，每天对入组患者的各项临床资料和生理生化指标进行收集和观察记录，包括病情轻重评分（APACHEⅡ评分，由急性生理评分、年龄评分和慢性健康状况评分3部分组成）、序贯器官衰竭评估（SOFA评分，包括呼吸、凝血、肝脏、循环、中枢神经和肾脏6个器官）和格拉斯哥（GCS）昏迷评分；检测体质指数、血乳酸水平和血糖水平；治疗和药物使用情况（是否使用机械通气、镇静剂、血管活性药物、肠内营养等）。将与AGI分级相关的临床因素和胃肠音指标（前期研究有5项胃肠音指标与AGI分级呈负相关，分别为胃部位测得的胃肠音波数量、胃肠音波时间百分比、胃肠音波最大时间、胃肠音波平均功率及胃肠音波最大功率）纳入分析，取主成分分析后的前5个主成分，通过BP网络模型训练，构建神经网络模型，从而建立危重患者AGI分级的综合预测模型。结果60例患者年龄19～98（平均54.6）岁，男性42例（70.0%）。其中多发骨折22例，严重感染15例，颈椎骨折和主动脉修复术后各7例，中毒后意识障碍5例，脑外伤4例；急诊手术33例，择期手术10例，内科治疗17例。有糖尿病史56例（93.3%）；使用血管活性药物45例（75.0%），机械通气者37例（61.7%），肠内营养44例（73.3%）。APACHEⅡ评分为4.0～28.0（平均16.8）分。与AGI分级呈显著正相关的有4项临床因素，乳酸水平（r = 0.215，P = 0.000）、SOFA评分（r = 0.383，P = 0.000）、血管活性药物（r = 0.611，P = 0.000）和机械通气使用与否（r = 0.142，P = 0.014），加上与AGI分级呈负相关的5项胃肠音指标，将这9个与AGI分级相关性较强的指标组成333 × 9的特征矩，进行主成分分析后，筛选出5个主成分进入BP神经网络模型，训练测试后得到危重患者AGI分级综合神经网络模型。模型的拟合度为0.967 3，对AGI分级预测的准确率为82.61%。结论根据胃肠音指标和相关临床影响因素构建的AGI分级综合预测模型，可为医护人员进行危重患者AGI分级的进一步预测判断提供参考。

ObjectiveTo develop the comprehensive prediction model of acute gastrointestinal injury （AGI） grades of critically ill patients.MethodsFrom April 2015 to November 2015, the binary channel gastrointestinal sounds （GIS） monitor system which has been developed and verified by the research group was used to gather and analyze the GIS of 60 consecutive critically ill patients who were admitted in Critical Care Medicine of Chinese PLA General Hospital. Also, the AGI grades （GrandeⅠ-Ⅳ, the higher the level, the heavier the gastrointestinal dysfunction） were evaluated. Meanwhile, the clinical data and physiological and biochemical indexes of included patients were collected and recorded daily, including illness severity score （APACHE Ⅱ score, consisting of the acute physiology score, age grade and chronic health evaluation） , sequential organ failure assessment （SOFA score, including respiration, coagulation, liver, cardioascular, central nervous system and kidney） and Glasgow coma scale （GCS） ; body mass index, blood lactate and glucose, and treatment details （including mechanical ventilation, sedatives, vasoactive drugs, enteral nutrition, etc.） Then principal component analysis was performed on the significantly correlated GIS （five indexes of gastrointestinal sounds were found to be negatively correlated with AGI grades, which included the number, percentage of time, mean power, maximum power and maximum time of GIS wave from the channel located at the stomach） and clinical factors after standardization. The top 5 post-normalized main components were selected for back-propagation （BP） neural network training, to establish comprehensive AGI grades models of critically ill patients based on the neural network model.ResultsThe 60 patients aged 19 to 98 （mean 54.6） years and included 42 males （70.0%） . There were 22 cases of multiple fractures, 15 cases of severe infection, 7 cases of cervical vertebral fracture, 7 cases of aortic repair, 5 cases of post-toxicosis and 4 cases of cerebral trauma. There were 33 emergency operation, 10 cases of elecoperectomy and 17 cases of drug treatment. There were 56 cases of diabetes （93.3%） . Forty-five cases （75.0%） used vasoactive drugs, 37 cases （61.7%） used mechanical ventilation and 44 cases （73.3%） used enteral nutrition. APACHE Ⅱ score were 4.0 to 28.0 （average 16.8） points. Four clinical factors were significantly positively related with AGI grades, including lactic acid level （r = 0.215, P = 0.000） , SOFA score （r = 0.383, P = 0.000） , the use of vascular active drugs （r = 0.611, P = 0.000） and mechanical ventilation （r = 0.142, P = 0.014） . In addition to the five indexes of gastric bowel sounds which were found to be negatively correlated with AGI grades, the characteristics of 333 by 9 were composed of these nine indexes with high correlation of AGI grades. Five main components were selected after principal component analysis of these nine correlated indexes. A comprehensive AGI grades model of critically ill patients with a fitting degree of 0.967 3 and an accuracy rate of 82.61% was built by BP artificial neural network.ConclusionThe comprehensive model to classify AGI grades with the GIS is developed, which can help further predicting the classification of AGI grades of critically ill patients.