Dynamic Prediction of Acute Graft-versus-Host-Disease with Longitudinal Biomarkers

摘要

This dissertation builds three prediction tools to dynamically predict the onset of acute graft-versus-host disease (aGVHD) with longitudinal biomarkers. Acute graft-versus-host disease is a complication for patients who have received allogeneic bone marrow transplant, and it is fatal for some patients. Clinicians could benefit from these prediction tools to identify patients who are at risk and who are not, and thus assign appropriate interventions.;Our first project introduces how to apply joint modeling with latent classes (JMLC) and landmark analysis to aGVHD data. In JMLC, we group all aGVHD-free patients into one latent class and define that class as the "cure" class. In landmark analysis, we incorporate patients' biomarker information up to the landmark time to gain more efficiency. Computer simulations show that both methods adjust for the measurement error, and that JMLC outperforms landmark analysis when the functional form of the biomarker profile is correctly specified.;In our second project, we describe how to execute dynamic prediction with the pattern mixture model, in which each patient is classified by his/her time-to-aGVHD, and patients in the same group share the same mean profile of biomarkers. The pattern mixture model is easy to execute and straightforward to interpret. Simulations indicate that the pattern mixture model controls loss of accuracy in predictions.;In our third project, we incorporate censored cases to generalize the pattern mixture model in the second project. The simulation results demonstrate that this generalized pattern mixture model accurately estimates of the marginal pattern probabilities, and thus better estimates early predictions compared to early predictions not incorporating censored observations.;In our fourth project, we explain the process of parametric bootstrap in selecting the number of latent classes in JMLC. Compared with the standard information-based criteria in model selection in JMLC, our parametric bootstrap likelihood ratio test (LRT) controls the Type I error well while maintaining sufficient power. We also propose two sequential early stopping rules to relieve the computational burden of bootstrap.