Dynamic Predictions of Visual Acuity in Uveitis


Uveitis is the inflammation (INF) of the uveal tract, i.e. the inner eye, and leads to reduction of visual acuity (VA). In contrast to other eye diseases that affect mostly elderly patients, uveitis affects young patients. Therefore, it is the leading cause of legal blindness in the working age population in the western world. Physicians utilize different types of information to predict patient VA outcome. This includes longitudinal data, such as INF, complication (COM) and other patients characteristics.

A popular framework to analyze outcomes measured over time is the multivariate mixed-effects model. We aim to obtain individualized dynamic predictions of VA using all available patient information and accounting for the special features of the data set. In particular, the first challenge is to take into account that patients can be categorized in sub-groups that exhibit different progression rates for high, blind and moderate VA. Furthermore, the longitudinal outcomes can be connected 1) via random effects and 2) via association parameters between them. A characteristic of longitudinal data, such as INF and COM, is that several features of these variables could be associated with VA (underlying value, slope, area under the curve). This could lead to a large number of parameters that need to be estimated. Therefore, the second challenge is to identify an appropriate set of predictors for the outcome of interest VA.

We propose a latent class multivariate mixed model to identify different sub-groups of patients and we compare two popular approaches that deal with variable/model selection: 1) a Bayesian shrinkage approach, where both INF and COM are included in the VA prediction model, and 2) Bayesian model averaging where INF and COM are included separately in the VA model. The two approaches are compared using a cross-validation procedure for model evaluation. Prediction accuracy is deemed acceptable when the difference between the predicted and the observed value is considered to have a non-detectable visual change for the patients. The motivation comes from a study of 365 uveitis patients with a mean age of 44. These patients visited the Rotterdam Eye Hospital in the period from 2000 to 2014.

International Biometric Conference
Barcelona, Spain
Eleni-Rosalina Andrinopoulou
Assistant professor in Biostatistics