Illness-death models are commonly used to study chronic diseases characterized by multiple stages, where subjects are at a non-negligible risk of death. Examples of such diseases include cancer, HIV, and Alzheimer's disease. When the disease status can only be determined by periodic assessments, the exact entry times to states are unknown. We aim to use the multistate data under intermittent observation along with high-dimension covariates, to jointly predict disease progression and death at a particular time horizon. We formulate an illness-death model and a penalized likelihood in settings where the disease processes are under intermittent observation and death is subject to right censoring. An innovative expectation-maximization (EM) algorithm is then developed which can flexibly incorporate different penalty functions and allows one to exploit existing packages. The method will be illustrated in the context of a biomedical study to jointly predict a non-fatal event and death.