Enhancing clinical Named Entity Recognition (NER) requires tackling domain-specific terminology and limited annotated data. This study investigates two strategies: (1) synonym-based augmentation using SNOMED CT ontology and (2) synthetic note generation leveraging GPT-4.1. We evaluate transformer models—BioBERT, RoBERTa, and DeBERTa-v3-large—on both performance and efficiency. It is found that DeBERTa-v3-large with synonym augmentation achieved the best results (F1=0.821), outperforming domain-specific and general-domain baselines. In contrast, GPT-based data showed mixed effects due to semantic drift. While augmentation slightly improved accuracy, it also increased training time, underscoring trade-offs relevant to inference efficiency. These findings offer insights for optimizing clinical term entity linking pipelines under computational constraints.