Dynamic Prediction of Clinical Events with Competing Risks Using Longitudinal Biomarker Data
Abstract
In many clinical investigations, a large amount of biomarker or vital sign data are collected repeatedly over a long-term follow-up until some terminal clinical event occurs. Researchers are interested in not only studying the association of such biomarkers with disease progression, but also making use of the resourceful information of the biomarkers and their change over time to predict the probability of future events. Our research context is the study of kidney disease, where predicting the risk of native kidney failure or kidney allograft failure is essential for making personalized treatment plans. We propose to develop novel risk prediction models using longitudinal data to 1) predict end-stage renal disease (ESRD) among patients with chronic kidney disease (CKD) and 2) predict renal graft failure among kidney transplantation (KTx) recipients. In order for clinicians to implement the model in medical practice, the method has to be simple yet flexible enough to handle different situations. Current methods of dynamic prediction using joint models of longitudinal and survival outcome are often computationally prohibitive in practical situations, therefore, we propose a statistical framework for dynamic prediction that is easy to interpret, simple to implement and applicable to a broad range of practical situations. In the first project, we adapt the dynamic prediction framework using the landmark approach [Van Houwelingen, 2007] and its extension Li et al. [2016], with application to predict post-transplant renal graft failure. Such a dynamic model incorporates post-transplant lab measures, non-terminal medical events, and medical history data to make real-time predictions. The model’s performance is compared with a Cox model developed at baseline (“static” model), which is then reapplied with time-updated biomarker data to make predictions. A comparison of model discrimination and calibration indicates that the dynamic model is capable of capturing the changes in risk factors, the at-risk population, the time-varying covariate effects and the baseline hazard overtime, and therefore is more suitable than the static model in the longitudinal context. Next, we extend the proposed dynamic prediction framework to the competing risk setting. Death is a common competing event for ESRD or renal graft failure; however, failure to account for such a competing risk is a common methodological problem in the studies currently available. In the presence of competing events, it is of primary interest to predict the event-specific cumulative incidence function (CIF). We propose a dynamic prediction framework by directly modeling the CIF of ESRD using the proportional subdistribution hazard model [Fine and Gray, 1999]. In the third project, we develop a nonparametric method to quantify and estimate the predictive accuracy of the model with competing risk data. We consider the time-dependent discrimination and calibration metrics, including the time-dependent receiver operating characteristic curve and the Brier score. We address censoring by weighting the censored subjects with the conditional probability of the event of interest given the observed data.
Subject Area
Biostatistics
Recommended Citation
Wu, Cai, "Dynamic Prediction of Clinical Events with Competing Risks Using Longitudinal Biomarker Data" (2017). Texas Medical Center Dissertations (via ProQuest). AAI10680428.
https://digitalcommons.library.tmc.edu/dissertations/AAI10680428