Adapting Pretrained Language Models for Solving Tabular Prediction Problems in the Electronic Health Record