Language Model Training Paradigms for Clinical Feature Embeddings