Protected Health Information De-Identification on Visual and Textual Features using Transfer Learning

Cyber data breaching has tremendously increased over the years, especially in countries like India. Medical documents have information more valuable than any other document shared online. Various laws came into existence which restricts the flow of protected health information in any sort of communication. Considering the possible threats involved in medical data transfer, there is a need to de-identify protected health information in any type of medical document effectively and also minimize the possibility of reverse engineering or re-identification. Existing works completely focus on electronic health records with only textual data. Our aim is to present the importance of de-identification in visual data along with textual features and also show a method for achieving it using a transfer learning based approach. An end-to-end artificial intelligence based application is presented which de-identifies all types of protected health information from either an image or a pdf with both textual and image data or DICOMs.