BRATsynthetic: Text De-identification using a Markov Chain Replacement Strategy for Surrogate Personal Identifying Information

Objective: Implement and assess personal health identifying information (PHI) substitution strategies and quantify their privacy preserving benefits. Materials and Methods: We implement and assess 3 different `Hiding in Plain Sight` (HIPS) strategies for PHI replacement including a standard Consistent replacement strategy, a Random replacement strategy and a novel Markov model-based strategy. We evaluate the privacy preserving benefits of these strategies on a synthetic PHI distribution and real clinical corpora from 2 different institutions using a range of false negative error rates (FNER). Results: Using FNER ranging from 0.1% to 5% PHI leakage at the document level could be reduced from 27.1% to 0.1% (0.1% FNER) and from 94.2% to 57.7% (5% FNER) utilizing the Markov chain strategy versus the Consistent strategy on a corpus containing a diverse set of notes from the University of Alabama at Birmingham (UAB). The Markov chain substitution strategy also consistently outperformed the Consistent and Random substitution strategies in a MIMIC corpus of discharge summaries and on a range of synthetic clinical PHI distributions. Discussion: We demonstrate that a Markov chain surrogate generation strategy substantially reduces the chance of inadvertent PHI release across a range of assumed PHI FNER and release our implementation `BRATsynthetic` on Github. Conclusion: The Markov chain replacement strategy allows for the release of larger de-identified corpora at the same risk level relative to corpora released using a consistent HIPS strategy.