De-identification algorithm for free-text nursing notes

All personally identifiable information must be removed from patient medical records before the data can be shared with other researchers. We present an automated method of removing protected health information (PHI) from free-text nursing notes taken from a U.S. hospital. We have previously shown that one clinician can locate PHI in nursing notes with an average sensitivity of 0.81, and for teams of two clinicians the sensitivity is 0.94. Our method uses lexical look-up tables, regular expressions, and simple heuristics to locate PHI with an overall sensitivity of 0.92 (0.98 for names, 0.96 for dates), which is significantly better than the average sensitivity of a single human. The algorithm has a positive predictive value of only 0.44, so additional software was developed to allow the user to review the terms identified as PHI and manually eliminate false positives. The algorithm is open-source and will be made freely available on PhysioNet together with a re-identified corpus of nursing notes