基于纵向研究中协变量依时理论与效应分层Lasso方法的老年人群2型糖尿病发病风险预测模型研究

Risk prediction is important for the prevention of Type 2 diabetes mellitus(T2DM). Most of the current risk prediction methods for T2DM are based on the assumption that the covariates are time-independent, but the prediction accuracy can be significantly degraded if some risk factors are highly dynamic in a longitudinal study. Most of studies on the interaction of risk factors for T2DM assume that main effect and interaction are on the same order, yet the performance of effect selection can be poor if there are many covariates in the study. The methods of statistical inference for the risk prediction model with time-dependent covariates and risk factor selection based on the effect hierarchy principle were proposed in our previous study. This project is based on the Beijing Longitudinal Study on the Elderly with 25-year follow-up period. Using time-dependent information including physical conditions, living habits and biochemical indexes should be an important consideration in the longitudinal study, and competing risk of the death should be a reasonable consideration in the elderly population. Therefore, the competing risks models with time-dependent covariates are proposed to describe the relation between the dynamic covariates and the risk of T2DM among the elderly in this project. Considering sparsity and hierarchy principles, the two steps method is used to reduce the dimension of the covariates in the longitudinal study. The hierarchical Lasso is presented to simultaneously select risk factors and interactions on the risk of T2DM in the elderly population. Comparisons among different risk assessment models and different effect selection methods are conducted by simulation studies and validation, and the best model is selected. The dynamic prediction model for the risk of T2DM among the elderly is obtained by using the Landmarking approach, and the prediction accuracy can be raised. This study proposes a new method for effect selection of risk factors, interaction analysis and risk prediction of T2DM and the similarly chronic non-communicable diseases in the elderly population.

风险预测对2型糖尿病（T2DM）预防非常重要。传统T2DM发病风险预测研究假定协变量不随时间改变，某些因素动态变化时预测精度会较低；因素交互作用研究一般假定协变量主效应和交互效应同阶，多因素分析时效应筛选性能会较差。课题组前期开展了依时风险评估与效应分层变量选择探索研究，本项目基于北京老年人群25年多维纵向数据，考虑身体状态、生活习惯和生化指标等多个因素依时性，以死亡作为T2DM发病竞争事件，建立依时竞争风险模型描述协变量动态变化与风险关系；考虑效应稀疏和分层准则，通过两步法对高维纵向数据降维，采用分层Lasso方法选择变量主效应和交互效应，准确筛选出老年人群T2DM发病风险因素与交互作用；通过随机模拟和随机抽样验证选择最优评估模型，采用Landmarking方法构建老年人群T2DM发病风险动态预测模型，提高预测精度。本研究将为T2DM等疾病发病风险因素与交互作用筛选和风险预测提供新方法。

2型糖尿病（T2DM）风险预测传统研究假定协变量不随时间改变，而纵向研究中个体的身体状态、行为习惯和生化指标等会发生动态变化，使得风险因素筛选准确度和效应估计精确度不高。本项目基于北京老龄化多维纵向研究25年随访数据,考虑相关因素作为依时协变量纳入评估模型，更加精细的描述了风险因素动态变化与发病风险的关系 。基于依时风险评估模型，采用两步法筛选出老年人群T2DM发病风险因素，包括年龄、吸烟量、身体质量指数（BMI）、甘油三酯（TG）和高密度脂蛋白胆固醇（HDL-C），采用分层Lasso方法，筛选得到TG和HDL-C存在交互作用。通过随机模拟，比较了依时Cox模型和基线Cox模型 在风险因素筛选和效应估计方面的性能，考虑协变量依时的Cox回归模型显著提高风险因素筛选准确性和效应估计的精确性。采用依时Cox回归模型和考虑死亡作为竞争事件的依时竞争风险模型（Fine-Gray model），评价了HDL-C、TG、TG/HDL-C比值、吸烟与T2DM风险的关联性，验证了相关风险因素与T2DM发病的关系。采用10折交叉验证的方法，评价比较了逐步递增TG/HDL-C比值、基线空腹血糖（FPG）下 3个层次模型的预测性能，发现年龄、吸烟量、BMI、TG、HDL-C、TG/HDL-C比值、基线FPG对T2DM发病风险预测性能最高，依时Cox模型和依时竞争风险模型对10年的T2DM发病风险预测指标-受试者工作特征曲线（receiver operating characteristic curve，ROC曲线）下面积（AUC）及95%置信区间（95% CI）分别达到0.743 (0.680-0.806)和0.755 (0.698-0.813)，确定T2DM发病风险最优预测模型纳入的变量为：年龄、吸烟量、BMI、TG、HDL-C、TG/HDL-C比值和基线FPG。基于得到的老年人群T2DM发病风险因素和之间的交互作用，采用Landmarking方法，估计得到风险因素主效应与交互效应，建立了老年人群T2DM发病风险动态预测模型，该模型预测性能指标C-index（95% CI）达到0.752 (0.733-0.770)，预测性能良好，为糖尿病等慢性非传染性疾病风险预测与预防提供了新模型新方法。

内容获取失败，请点击重试