Risk prediction plays an important role in precision medicine. In many clinical settings, it is of great interest to develop models for predicting the t-year risk of developing a clinical event using baseline covariates. Such t-year risk models can be estimated by fitting a flexible time specific generalized linear model (GLM). However, efficient and robust estimation of the risk model is challenging under heavy censoring. Incorporating intermediate outcome information could potentially improve the efficiency of the prediction model. However, existing augmentation methods largely do not allow intermediate outcomes to be subject to censoring and may yield invalid results under model mis-specification. In this paper, we propose a two-step augmentation method to improve the estimation of t-year risk model by leveraging longitudinally collected intermediate outcome information that is subject to censoring. We also propose resampling methods to assess the variability of our proposed estimators. Numerical studies show that the proposed procedures perform well in finite sample. We also illustrate the proposed methods using data from Prevention Diabetes Program.