AUTOMATED DEIDENTIFICATION OF PATHOLOGY AND RADIOLOGY DATA

The overall project objective is to optimize and automate the detection and redaction of PHI in radiology data (DICOM) and pathology data (Whole Slide Imaging - WSI) – both header metadata and “burned in” pixel data – thereby minimizing the need for human-in-the-loop review/manual redaction and maximizing deidentified data throughput. Impact Business Information Solutions, Inc (IBIS) has established four major goals for Phase II of this project.

One, to enhance the deep learning models begun in Phase I, executing the algorithms against real-world data (as opposed to synthetic data, in Phase I), and incrementally improving their success rates. IBIS will also enhance the NLP capability for PHI detection in text with a more sophisticated algorithm. Two, to put in place a comprehensive

human-in-the-loop workflow. This will require the addition of several new modules to EICON DEID including User Management, Access Control, and a Messaging subsystem for event-driven notifications. Three, to develop a more accurate automated measurement of de-identification confidence using an uncertainty quantification algorithm, which will be used to determine with informed precision the trigger point for initiation of the human-in-the-loop process. And four, to productize and validate the EICON DEID solution with a

strong focus on system performance, scalability, usability and regulatory compliance.