Many health-related studies aim to explore how time-to-death is associated with changes in health status. Noting that health status is an internal time-varying covariate, likelihood-based procedures are thus inapplicable in this application. With the objective of providing risk predictions based on one's current health status, this presentation discusses strategies to obtain such predictions. We summarize the incomplete internal covariate process with latent variable(s), and adopt a joint modelling framework that links the survival and health outcome processes together. The proposed estimation procedure extends the conditional score approach of Tsiatis and Davidian (2001) by allowing the longitudinal sub-model(s) to accommodate correlated successive observations. We motivate and illustrate the proposed modelling with a dataset from administrative health records, and demonstrate adequate predictive performance through a simulation study. This is joint work with X. Joan Hu (SFU).