Measuring the gap: correlating synthetic-to-real drift with PHI de-identification performance

Clinical text de-identification enables the use of electronic health records while protecting patient privacy, but public training data remain scarce and often have mismatched documentation styles. Recent works have proposed using large language models (LLMs) to generate synthetic clinical notes, but it remains unclear if they reflect distributions of real clinical notes. We examine how lexical and semantic drift across training and evaluation corpora affects protected health information (PHI) tagger performance. We generated synthetic notes from scratch for four categories using five generator LLMs and one judge LLM. Next, we fine-tuned small de-identification models on real, synthetic, and mixed corpora, and evaluated them on three external benchmarks under a harmonized label schema. Models trained on broad, clinically oriented sources transfer better than those on legal or narrowly synthetic data. These results suggest that although synthetic data lacks some real-world distributional properties, it remains useful in low-resource settings. We found that compact distributional and embedding-based drift measures moderately correlate with out-of-distribution F1 score, a practically important result because drift estimation can improve synthetic-data quality control and alignment.