Anonymization of Electronic Health Care Records: The EHR Anonymizer

Abstract

Electronic health care records (EHRs) are the main source of information between medical professionals. Therefore, EHRs would be a good source for advanced data analysis. However, the current data privacy laws complicate the utilization of patient data; due to the concern of evaluating personal information without consent. Since personal information is only partially useful for analysis, removing it from the documents is a reasonable solution that respects the patients privacy and makes patient data accessible for third parties. The electronic health care records anonymizer (EHR) is a BRAT-based application that can be used to tag personal information in EHRs and replace them with previously defined placeholders. In a first annotation step, machine learning (Stanford NER) and rule based approaches generate tags for potential personal information and prior user annotations are added to the document. The user can now add/edit/remove annotations until the annotations are adjusted. The performance of the automated tagging was evaluated on the basis of 10 example EHR input documents, and resulted in a F1 score of 0.72 and an accuracy of 0.96. Since the goal was to keep false negative (FN) tags at a low rate, the false negative rate (FNR) of 0.1 motivated the implementation of a ‘feedback loop’ that saves user annotations for future uses and therefore reduces FNs. Further, the moderate number of false positive (FP) tags results a low precision score of 0.6, which could be counteracted with the implementation of a ‘black list’ that saves tags that were deleted by the user. In combination the ‘feedback loop’ and ‘black list’ would boost the automatic annotation performance, yet in the end, the user has control.