README.md
In JanvandenBrand/highdimjm: High Dimensional Joint Models

highdimjm

Project supported by the Dutch Kidney Foundation (www.nierstichting.nl) Grant NSN19OK003

Project outline

Personalized medicine after kidney transplantation requires accurate forecasting of graft function. Repeated measurements of renal function during follow-up are therefore crucial. Graft functional assessment guides the decision to perform biopsies and is used for prediction of outcome. Yet, in routine practice, the trajectories of graft function over time after transplant are interpreted without validated statistical tools to assess those trajectories. At present, prediction models incorporate only a snapshot of cross-sectional data, usually those taken at the time of transplantation or a year after transplantation (in case of prognostic markers), or around the time of a suspected rejection episode (in case of diagnostic markers). The statistical models used to make predictions do not take changes in biomarker values over time into account, and therefore lack clinical utility.

I will use a cutting edge multivariate joint modeling approach to create a dynamic prediction model that can forecast graft function, acute rejection, graft failure, or death with a functioning graft in patients who received a kidney transplantation.

In work package 1, I will build upon my previous work on univariate joint models. In summary, the high-dimensional joint model is fitted in three stages: 1. all possible pair wise combinations of univariate mixed models are created, 2. these are joined into a multivariate mixed model, and 3. the multivariate, mixed models are joined to a survival outcome using a discriminant analysis approach.

The last step distinguishes my approach from other implementations of the multivariate joint model. It enables computation so efficiently that it can be done with standard consumer hardware. Furthermore, I will go beyond the state-of-the-art and extend the high-dimensional multivariate joint model to multiple events. The event-specific multivariate mixed models are joined to arrive at a high-dimensional multivariate joint model. The validity of the approach will be tested by comparing results to the previous work. Computational performance will be benchmarked against other approaches such as JMBayes and joineRML. In work package 2, I will use the technology developed in WP1 to create a dynamic prediction model that uses data routinely collected in clinical practice to predict graft function, acute rejection, graft failure, and death with a functioning graft. I will use data available from the Biobank Renal Transplantation at KU Leuven. In this extremely well annotated cohort, data on transplant function are available, exactly at time of the biopsies and also collected routinely in between biopsies. The database also includes all relevant donor and recipient demographic variables, and outcome parameters. Predictive performance will be quantified using standard measures including the C-index, R2, and calibration between predicted and observed risks at various time points. To externally validate the model, I will use a cohort of similar patients selected from Radboudumc.

Functions to support model fitting and reporting.