Early Risk Prediction with Temporally and Contextually Grounded Clinical Language Processing