利用联合模型动态预测IgA肾病患者进入终末期肾病的概率

Dynamic prediction of end stage kidney disease for patients with IgA nephropathy

目的:建立一个工具能够根据每次访视IgA肾病(IgAN)患者的病情变化,动态预测患者进入终末期肾病(ESKD)的概率。方法:选取1997—2007年在国家肾脏疾病临床医学研究中心长期随访的IgAN患者。利用联合模型(Joint model)分析IgAN患者所有访视点的产生纵向数据,实现动态预测患者ESKD的概率,以达到个体化动态预测的目标。我们首先以肾活检时刻为基准建立了临床模型(模型A)和临床病理模型(模型B)。模型A的生存子模型部分引入了年龄、性别、尿蛋白定量(Log转换)、估算的肾小球滤过率(eGFR)、eGFR下降斜率、血白蛋白水平等指标;纵向子模型部分引入了年龄、尿蛋白定量、eGFR、eGFR下降斜率以及血清白蛋白水平。模型B在模型A的基础上生存子模型上增加了牛津病理分类指标。模型C是以首次访视时间为基准重新拟合了模型A并评估了其性能。所有的纵向子模型采用访视时间t的非线性混合模型拟合。结果:本研究纳入了866例IgAN患者,以首次门诊就诊为基线,平均随访时间为12.2±5.5年,共访视22 533人次,平均每个患者访视26次。随访中260例(30.0%)患者进入ESKD,其末次随访平均eGFR为9.1±3.0 mL/(min·1.73m^(2))。将所有患者按照3∶1的比例,随机分为建模组(650例)和测试组(216例)。以肾活检时刻为基线,模型A与模型B的性能基本一致,二者均表现出较高的预测性能,纳入或者剔除病理学变量并未明显增加联合模型对ESKD风险预测的准确性。随着随访时间的增加,模型A的预测性能持续提升,在肾活检后第5年左右达到最佳性能。AUC值由肾活检时的0.864增至肾活检后第5年的0.956;Brier评分由肾活检时的0.124降至活检后第5年的0.058。以首次访视时刻为基线的模型C也取得了类似的结果。为方便临床实践,我们利用Shiny包实现了动态预测,并将相关模型R对象、源代码公布在网络上。结论:联合模型可以用于IgAN患者访视点产生的纵向数据,高性能动态预测IgAN患者尿毒症的概率,实现个体化动态预测的目标。

Objective:To develop a personalized dynamically model to predict the risk of kidney failure for patients with IgA nephropathy(IgAN)using updates of longitudinal data at each follow⁃up visit.Methodology:Three joint models were fitted to analyze the longitudinal data at each visit.We defined the baseline as the time of the kidney biopsy and fitted a clinical joint Model A,which included variables such as sex,age,eGFR,ALB,and proteinuria.We also constructed a clinical⁃pathological Model B,which incorporated both clinical and histological MEST features.Model C was fitted using parameters identical to those in Model A,however,the baseline was defined as the time of the clinical visit instead of the biopsy.Results:A total of 866 patients were included(650 in the development cohort and 216 in the validation cohort)and contributed 10565 patient⁃years of data,and 22533 eGFR and proteinuria measurements.Models A and B performed similarly with high predictive ability.However,the inclusion or exclusion of pathological variables did not significantly increase or decrease the accuracy of the joint models for predicting kidney failure risk.As follow⁃up time increased,model A's predictive performance continued to improve,reaching optimal performance around 5 years after kidney biopsy.The AUC value increased from 0.864 at kidney biopsy to 0.956 at 5 years after biopsy,and the Brier score decreased from 0.124 at the time of biopsy to 0.058 at 5 years after biopsy.The model C achieved similar results.All predictive performances were confirmed in the validation cohort.To facilitate clinical practice,we utilized the Shiny package to implement dynamic prediction.The R objects and source code have been made publicly available online.Conclusion:The Joint model can be utilized for the longitudinal data generated from visits of IgAN patients,providing a high⁃performance dynamic prediction of kidney failure for IgAN patients.This helps achieve the objective of individualized dynamic prediction.