With growing interest in personalized medicine and the rise of machine learning, building good risk prediction and prognostic models has been drawing renewed attention. In this development, much effort is often concentrated in identifying good predictors of patient outcomes. However, the same level of rigor is often absent in improving the outcome side of prediction. This study focuses on this crucial but neglected aspect of model development. Using latent variable (LV) strategies is a promising way of constructing valid and reliable prediction targets utilizing rich patient outcome data. Their flexible nature is a big advantage in organizing and synthesizing complex multivariate outcomes. However, the same nature can also make LV solutions esoteric and subjective. As a pragmatic solution to this problem, we propose a learning framework that focuses on the use of explicit clinical validators, which bridge exploratory LV solutions and concrete clinical intentions. We will show an application to a longitudinal study of manic symptoms, where predicting risk trajectory types is crucial in treatment decisions and patient care.