Privacy-Preserving Generation of Synthetic Pathology Reports for Information Extraction
Résumé
A long-standing goal in clinical NLP is extracting key variables from clinical text, but progress is hindered by domain shift, limited annotated data, and privacy constraints. In this paper, we propose a novel privacy-preserving method to generate synthetic report/data pairs that support information extraction by associating LLM-generated pathology reports with thirteen variables commonly found in real reports for breast cancer patients. We first generate synthetic tabular data for these variables and their possible values, comparing several tabular synthesizers and selecting PATE-CTGAN for its strong statistical fidelity and differential privacy guarantees. We then generate pathology reports using three LLMs to maximize linguistic diversity and conditioning generation on synthetic variable-value sets. We show that a model fine-tuned on the resulting synthetic data substantially outperforms the same model without fine-tuning and improves performance over a 3-shot baseline using synthetic in-context examples, achieving 0.79 accuracy compared to 0.38 and 0.64, respectively. These results show that high-quality synthetic data can effectively compensate for limited annotated clinical data while enabling accurate and privacy-preserving clinical information extraction. The code is available at https://gitlab.inria.fr/aabdelnu/syntheticpreports.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| licence |
