Figure 5.

Our rule-based synthetic PHI algorithm incorporates modules for each PHI category that can easily be augmented with new modules that handle new types of PHI. For each PHI category, it relies on a parser, a constraint, and a generator. The parser infers the input distributions of content and format. These distributions are skewed by constraints that, for instance, prevent generation of a PHI token identical to the original PHI and maintain the relative order and format of dates in each report. Finally, following these distributions, a rule-based generator replaces each true PHI span with a synthetic one. We elected not to use a neural-based generator to avoid the risk of outputting training data. This figure only contains synthetic PHI. PHI: protected health information.