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

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Machine Learning

Background:

• Electronic health records (EHRs) require de-identification for patient privacy.
• Publicly available de-identification training data are scarce and often lack consistent documentation styles.
• Large language models (LLMs) offer a potential solution for generating synthetic clinical notes.

Purpose of the Study:

• To evaluate the impact of lexical and semantic drift on protected health information (PHI) tagger performance.
• To assess the utility of LLM-generated synthetic clinical notes for training de-identification models.
• To determine if synthetic data reflects real-world clinical note distributions.

Main Methods:

• Generated synthetic clinical notes using five generator LLMs and one judge LLM.
• Fine-tuned de-identification models on real, synthetic, and mixed corpora.
• Evaluated model performance on three external benchmarks using a harmonized label schema.
• Assessed correlation between drift measures and out-of-distribution F1 score.

Main Results:

• Models trained on broad, clinically relevant data sources outperformed those trained on legal or narrowly synthetic data.
• Synthetic data, despite lacking some real-world distributional properties, proved useful in low-resource scenarios.
• Compact distributional and embedding-based drift measures showed moderate correlation with out-of-distribution F1 score.

Conclusions:

• LLM-generated synthetic clinical notes can augment scarce real-world data for de-identification tasks.
• Careful selection of training data sources and quality control through drift estimation are crucial for maximizing synthetic data utility.
• Drift estimation offers a practical method for improving synthetic data quality and alignment in clinical text de-identification.