2/17/2017

Speaker: Xuelin Huang, Department of Biostatistics, The University of Texas MD Anderson Cancer Center

[abstract]

Dynamic prediction is an important statistical tool for aiding medical decision-making, such as early detection of disease onset, and post-treatment monitoring of disease prognosis. For this purpose, subjects' biomarker values are repeatedly measured over time during follow-up visits. Predictions are conducted on a real-time basis so that at any time during follow-up, as soon as a new biomarker value is obtained, the prediction can be updated immediately to reflect the latest prognosis. Longitudinal biomarker trajectories are usually not linear, or even not monotone, and vary greatly across individuals. Therefore, it is difficult to fit them by parametric models. In this talk, I will first review the commonly used approaches for dynamic prediction, such as landmark analysis and joint modeling of longitudinal and survival data. Then I will introduce some of my recent work with colleagues to enrich the current methods. These include quantile regression on residual survival time, functional principal component analysis for summarizing the changing patterns of patients’ longitudinal biomarker trajectories, and a supermodel to smoothly extend landmark analyses on discrete time points to the whole follow-up time interval. Simulation studies show that the proposed approaches achieve stable estimation of biomarker effects over time, and are robust against model misspecification. Moreover, they have better predictive performance than current methods, as evaluated by the root mean square error and area under the curve of receiver’s operating characteristics. The proposed methods are applied to a data set of patients with chronic myeloid leukemia. At any time following their treatment with tyrosine kinase inhibitors, longitudinally measured transcript levels of the oncogene BCR-ABL are used to predict patients’ risks of disease progression.