基于核磁共振代谢组学特征构建脓毒症患者短期死亡风险的预测模型及效能

Construction of a predictive model of short-term death risk in patients with sepsis based on NMR metabolomics characteristics and its efficacy

目的 探讨基于核磁共振代谢组学特征构建脓毒症患者短期死亡风险的预测模型的效果。方法 选择2020年1月~2022年12月我院ICU收治的60例脓毒症患者作为训练集。以治疗后28 d时患者存活情况为分组标准,将所有患者分为存活组及死亡组。收集患者一般资料、实验室检查资料,在患者入组时搜集血清样本进行质子核磁共振特异性代谢标志物分析,通过LASSO回归筛选出排名前4的产物构建脓毒症患者短期死亡风险预测模型。最后纳入2023年1~12月于我院ICU治疗的脓毒症患者49例作为测试集对模型的预测效果进行验证。结果 死亡组患者APACHE Ⅱ评分高于存活组,差异具有统计学意义(P<0.05)。多元统计分析方法显示原始模型的预测能力大于任何1次随机排列Y变量的预测能力,证明模型有效。无监督主成分分析得分散点图共解释了57%的变量(PC1=50%,PC2=7%),脓毒症患者经PCPA造模后血清中内源性代谢物发生了显著变化。OPLS-DA图示模型拟合效果好,不存在特异点,两组的分布区域完全分开。载荷矩阵图显示L-天门冬氨酸、吲哚乙酸、丙氨酸等水平升高,异亮氨酸、亮氨酸等水平下降。提取OPLS-DA模型中VIP值最大的前50个变量进行非参数检验,最后得到34个有统计学意义的变量(P<0.05),共有19个与脓毒症死亡最有可能相关的特征代谢物。死亡组患者血清苯丙氨酸、肌酸、乙酰乙酸、谷氨酸、蛋氨酸、尿素、乳酸及氧化三甲胺显著高于存活组(P<0.05)。使用LASSO筛选出的4个高相关的代谢产物,并建立的列线图模型脓毒症患者短期死亡风险的C-index分别为0.993(95%CI:0.931~0.999)。列线图模型预测训练集的脓毒症患者短期死亡风险(AUC=0.993,95%CI:0.978~1.000,P<0.001)具有一定价值;将列线图模型预测测试集的脓毒症患者短期死亡风险(AUC=0.934,95%CI:0.863~1.000,P<0.001)具有一定的使用价值。模型对训练集预测的准确率分别为96.66%,敏感度分别为94.11%,特异度分别为97.60%。模型对验证集预测的准确率分别为90.62%,敏感度分别为88.23%,特异度分别为91.48%。结论 基于血清核磁共振代谢组学特征构建脓毒症患者短期死亡风险的预测模型对于脓毒症患者28 d死亡风险具有较好的预测价值。

Objective To investigate the effect of building a predictive model of short-term mortality risk in sepsis patients based on NMR metabolomics characteristics.Methods Sixty patients with sepsis admitted to ICU of our hospital from January 2020 to December 2022 were selected as the study objects.All patients were divided into survival group and death group based on the survival status at 28 d after treatment.General data and laboratory examination data of patients were collected.Serum samples were collected for proton nuclear magnetic resonance specific metabolic markers when patients were enrolled.The top 4 products were selected by LASSO regression to construct a short-term death risk prediction model for sepsis patients.Finally,patients with sepsis treated in ICU of our hospital from January 2023 to December 2023 were included to verify the predictive effect of the model.Results APACHE II scores in the death group were higher than those in the survival group,and the difference was statistically significant(P<0.05).Multivariate statistical analysis method shows that the prediction ability of the original model is greater than that of any one random arrangement of Y variables,which proves that the model is effective.The scatter plot of unsupervised PCA explained 57%of the variables(PC1=50%,PC2=7%),and the metabolic profile was different between the control group and the model group,with significant changes in serum endogenous metabolites in sepsis patients after PCPA modeling.OPLS-DA graphical model has good fitting effect,there are no special points,and the distribution areas of the two groups are completely separated.As shown in the loading matrix,the levels of L-aspartate,indoleacetic acid and alanine increased,while the levels of isoleucine and leucine decreased.The top 50 variables with the largest VIP value in the OPLS-DA model were extracted for non-parametric test.Finally,34 variables with statistical significance were obtained(P<0.05),and a total of 19 characteristic metabolites were most likely to be associated with sepsis death.Serum phenylalanine,creatine,acetoacetic acid,glutamic acid,methionine,urea,lactic acid and trimethylamine oxide in death group were significantly higher than those in survival group(P<0.05).The C-index of short-term death risk in sepsis patients was 0.993(95%CI:0.931-0.999)in an eriograms of four highly correlated metabolites selected by LASSO.The Nomogram model was of certain value in predicting the short-term mortality risk of sepsis patients(AUC=0.993,95%CI:0.978-1.000,P<0.001).The short-term mortality risk in model-validated sepsis patients with a Nomogram model(AUC=0.934,95%CI:0.863-1.000,P<0.001)was useful.The accuracy,sensitivity and specificity of the model to the training set were 96.66%,94.11%and 97.60%,respectively.The accuracy,sensitivity and specificity of the model to the validation set were 90.62%,88.23%and 91.48%,respectively.Conclusion The prediction model of short-term death risk in sepsis patients based on serum NMR metabolomic characteristics has a good value in predicting the 28 d death risk of sepsis patients.