基于机器学习的胰腺及上消化道肿瘤手术精细化营养支持的策略研究

Study on refined nutritional support strategy for pancreatic and upper gastrointestinal tumor surgery based on machine learning

目的通过机器学习方法建立模型探讨胰腺及上消化道肿瘤手术精细化营养支持的策略。方法回顾性分析2016年12月至2018年5月间北京协和医院基本外科120例行胰腺及上消化道肿瘤手术患者的临床资料。纳入患者年龄、身高、体重、体温、脉搏、呼吸频率、血红蛋白、白蛋白、血糖、能量摄入、蛋白质摄入等11项临床指标,以术后有无并发症为结局,采用支持向量机(SVM)交叉验证方法分别建立患者术后1、2、3 d并发症预测模型,绘制受试者工作特征曲线(ROC),计算曲线下面积(AUC)。纳入除蛋白质和能量摄入之外的9个变量进行SVM的逆运算,获取并发症发生率最低时的最适能量和蛋白质摄入区间。对其中50例患者的血清样品进行核磁共振氢谱(1H-NMR)波谱检测。运用偏最小二乘判别分析法(PLS-DA)获取患者术前1 d与术后1 d、术后1 d与术后3 d体内代谢物的变化,从人代谢组数据库(HMDB)获取相应的代谢物及代谢通路。结果89例患者中52例术后发生并发症,无并发症组和并发症组患者各临床指标的差异均无统计学意义。患者术后1、2、3 d SVM并发症预测模型的AUC值分别为0.80、0.65和0.72,最适能量区间分别为20.7~41.5、103.7~124.4、112.0~132.7 kJ·kg^(-1)·d^(-1),最适蛋白质区间分别为0.2~0.5、0.5~1.0、0.8~1.1 g·kg^(-1)·d^(-1)。基于1H-NMR代谢组学的PLS-DA模型得分图显示手术前后不同时间血液代谢产物出现显著变化。PLS-DA权重图离群值及HMDB筛选分析显示,术后1 d主要表现为糖异生增加,术后3 d主要表现为蛋白质分解增加。结论通过机器学习建模有助于获取胰腺及上消化道肿瘤手术后不同时点患者的最佳能量与蛋白质需求量区间,以此为依据可为适宜的个体化营养治疗提供方法学保证。

Objective To explore the refined nutritional support strategy for pancreatic and upper gastrointestinal tract tumor surgery through machine learning method.Methods Clinical data of 120 cases with pancreatic and upper gastrointestinal tract tumor who underwent surgery in the Basic Surgery Department of Peking Union Medical College Hospital between December 2016 and May 2018 were retrospectively analyzed.Eleven clinical indicators were recorded,including patients'age,height,weight,temperature,pulse,respiratory rate,hemoglobin,albumin,blood glucose,energy intake,protein intake,etc.The 1st,2nd,and 3rd postoperative complication prediction model were established,using support vector machine(SVM)cross-over validation method oriented to clinical outcome of complication.Receiver operating characteristic curves(ROC)were drawn and area under the curve(AUC)was calculated.Nine variables other than protein and energy intake were included in the inverse operation of SVM to obtain the optimal energy and protein intake intervals when the complication rate was lowest.Nuclear magnetic resonance hydrogen spectroscopy(1H-NMR)spectroscopy was performed on serum samples from 50 of the patients.Partial least squares discriminant analysis(PLS-DA)was used to obtain the changes of metabolites of the patients at 1 d preoperatively versus 1 d postoperatively and at 1 d postoperatively versus 3 d postoperatively,and the corresponding metabolites and metabolic pathways were obtained from the human metabolome database(HMDB).Results Postoperative complications occurred in 52 out of 89 patients,and the differences on each clinical index between the patients in the no-complication group and the complication group were not statistically significant.The AUC values of the SVM complication prediction model for patients at 1,2,and 3 d postoperatively were 0.80,0.65,and 0.72.The optimal energy intervals were 20.7-41.5,103.7-124.4,and 112.0-132.7 kJ·kg^(-1)·d^(-1),respectively,and the optimal protein intervals were 0.2-0.5,0.5-1.0,and 0.8-1.1 g·kg^(-1)·d^(-1).1H-NMR metabolomics-based PLS-DA model score plots showed significant changes in blood metabolites at different times before and after surgery.PLS-DA weight plots and HMDB screening analyses showed that increased gluconeogenesis was the main manifestation at 1 d postoperatively,and increased proteolysis was the main manifestation at 3 d postoperatively.Conclusions Machine learning modeling helps to obtain the optimal energy and protein requirement intervals for patients at different time points after pancreatic and upper gastrointestinal tumor surgery,which can provide methodological assurance for appropriate individualized nutritional therapy.