Study Confirms PHI Detectability Maintained with Surrogate Substitution.

This study investigates the effectiveness of structure-preserving de-identification, a technique that replaces sensitive protected health information (PHI) with realistic, same-type surrogates rather than generic placeholders. The goal is to maintain the fluency of clinical text and ensure compatibility with existing downstream analytical tools. The core question addressed is whether these substitutions corrupt the signal that PHI detectors rely on. Researchers developed a multi-detector evaluation protocol, focusing specifically on the masked spans where substitutions occur. Using equivalence testing across 11 detectors, 7 benchmarks, and 7 languages, they found that the recall on masked spans changed minimally, from 76.1% to 74.9%. This change was statistically equivalent to zero within a small margin, indicating that PHI detectability is largely preserved. Any minor loss in detection was attributed to malformed or out-of-distribution surrogates, rather than a fundamental degradation in detector performance. The study provides an open-source evaluation platform to audit structure-preserving transformations, confirming that well-formed substitutions are effective.