Walker Broadbent
Pronouns: he/him
Research Mentor(s): VG Vinod Vydiswaran
Co-Presenter: Weissman, Noah
Research Mentor School/College/Department: University of Michigan / Medicine
Presentation Date: April 20
Presentation Type: Poster
Session: Session 6 – 4:40pm – 5:30 pm
Room: League Ballroom
Authors: Walker Broadbent, Noah Weissman, Vinod Vydiswaran
Presenter: 70
Abstract
Michigan Medicine is home to many health documents, but they’re filled with information that could potentially be used to identify patients. Under HIPAA regulations, these documents are off limits to researchers, who could use these documents to identify trends and make breakthroughs in health research. Therefore, by developing an AI model that can recognize personal information and effectively “de-identify†them, the documents can be released to researchers. De-identification guidelines will guide the model through the training, which are being fine-tuned before the model is trained on hand de-identified data. The model will tag words and phrases under specific labels, such as age or name, and run them through an aliasing component, which replaces any tagged element with a specific replacement based on a different set of guidelines. The model must be able to identify multiple ways to write the same information, overcome spelling errors, and then return the aliased information in the same format. With the completion of this project, Michigan researchers will have access to health documents that will supplement their research endeavors. However, the model must first be evaluated before the product can be released: What specific patterns have slipped past the model? Are certain phrases being tagged incorrectly? Does the aliasing system correctly replace the tagged phrases?
Biomedical Sciences
