The SURROGATOR Framework for Context-Aware Surrogation of Privacy Sensitive Information in Medical Text

Clinical text cannot be shared freely due to data protection regulations. This paper introduces SURROGATOR, an open-source framework that replaces personal identifiable information with fictitious, high-fidelity synthetic surrogates to maintain text utility for human and artificial intelligence. We focus on the automated replacement of names, dates, and locations. Evaluation on a synthetic clinical corpus demonstrated high data utility, with named entity recognition performance remaining en par with the original text (F1-score of 0.70 vs. 0.73). Furthermore, a re-identification attack using a large language model resulted in an accuracy of 50.8%, which is equivalent to random chance. We conclude that SURROGATOR effectively balances patient privacy and data utility, providing a robust solution for the secure sharing of clinical documents in medical research.