Prediction from longitudinal high dimensional molecular data
Research Objectives:
Besides high dimensional molecular data available at baseline, repeated measurements of biomarkers of disease progression and/or response to treatment as time proceeds are often now routinely available, e.g., longitudinal gene expression predicting survival of burn patients, recovery of trauma patients or rejection in cardiac allograft recipients. Emphasis in joint modelling of longitudinal biomarkers and time-to-event data has been put on describing relationships and fitting sophisticated statistical models. Less attention has so far been given to practical use of the models for deriving dynamic predictions in individual patients.
Description of work:
In such longitudinal collections of high-dimensional molecular data, the extraction of important features is important and challenging. The use of a boosting approach will be investigated for this task that has already been successfully implemented for high-dimensional baseline data. Quantification of predictive performance is a central issue that will be addressed by using adapted prediction error curves, based on the Brier score for time-to-event data with timedependent covariates. The ultimate goal is a biologically meaningful joint model of highdimensional longitudinal and time-to-event data that also has practical relevance in the dynamic prediction of disease progression and/or response to a particular treatment.
Host Institution: University Medical Center Freiburg