基于贝叶斯多水平模型的平均血糖构建与应用

Average blood glucose construction: a multi-level Bayes model approach and application

目的探究基于贝叶斯多水平模型框架下平均血糖构建的方法,并对实例应用进行评价。方法通过贝叶斯多水平模型产生血糖的模拟数据,运用贝叶斯多水平模型3种方法同时构建平均血糖,并进行比较。采用队列研究方法,选择苏州某社区2011-2018年参加并随访的12321名45岁以上未发现卒中病史的成年居民作为研究对象,其中男性占比53.7%。利用完全贝叶斯方法计算出平均血糖并分为6组。运用Cox回归模型分析平均血糖值对致死性脑卒中发病的影响。结果 1000次模拟结果显示,用完全贝叶斯法计算的平均血糖估计误差均方为0.278,平均血糖估计值的平均误差为0.527 mmol,平均血糖与实际血糖相关系数平均值r=0.898。随访期间,人群中致死性卒中共153例,且平均血糖与致死性脑卒中发病风险有统计学意义(P<0.05),调整影响后,平均血糖大于140 mg/dL的发病风险是90~99 mg/dL的2.304倍(HR=2.304,95%CI 1.151~4.613)。结论完全贝叶斯多水平潜变量模型能准确估计平均血糖。

OBJECTIVE To explore the average blood glucose construction method based on the multi-level Bayes model and evaluate the example application. METHODS We generate simulated data with multi-level Bayes model. Three methods were utilized to construct the average blood glucose at the same time, then we compared the result with each other. A cohort study method was used to select 12321 participants aged over 45 y who without stroke in a community in Suzhou and was followed up from 2011 to 2018, of which 53.7% were male. Mean blood glucose calculated by the most accurate complete Bayesian method was divided into six groups. The Cox regression model was used to analyze the effect of mean blood glucose on the incidence of fatal stroke. RESULTS 1000 times of simulation result showed that the average mean blood glucose estimation calculated by the complete Bayesian method was 0.278, the average of blood glucose estimation was 0.527 mmol/L, and the average correlation coefficient with the actual blood glucose was r=0.898. During the follow-up period, 153 fatal strokes occurred. Association was found between the mean blood glucose and the risk of fatal stroke(P<0.05). The average risk of blood glucose over 140 mg/dL was 2.304 times that of 90-99 mg/dL(HR=2.304,95%CI 1.151-4.613) after the adjustment of effects. CONCLUSION The complete Bayesian multi-level latent variable model can accurately estimate the average blood glucose.