De-identification of Protected Health Information in Clinical Document Images using Deep Learning and Pattern Matching

Clinical documents that include lab results, discharge summaries, and radiology reports of patients are generally used by doctors for diagnosis and treatment. However, with the popularization of AI in healthcare, clinical documents are also widely used by researchers for disease diagnosis, prediction, and developing schemes for quality healthcare delivery. Though huge volumes of clinical documents are produced in various hospitals every day, they are not shared with researchers for study purposes due to the sensitive nature of health records. Before sharing these documents, they must be de-identified, or the protected health information (PHI) should be removed for the purpose of preserving the patient's privacy. If the documents are stored digitally, this PHI can be easily identified and removed, but finding and extracting PHI from old clinical documents that are scanned and stored as images or other formats is quite a daunting task for which machine learning models have to be trained with a large number of such images. This work introduces a novel combination of deep learning and pattern matching algorithms for the efficient de-identification of scanned clinical documents, distinguishing it from previous methods, which can primarily work only on text documents and not on scanned clinical documents or images. Thus, a comprehensive de-identification technique for automatically extracting protected health information (PHI) from scanned images of clinical documents is proposed. For experimental purposes, we created a synthetic dataset of 700 clinical document images obtained from various patients across multiple hospitals. The de-identification framework comprises two phases: (1) Training of YoloV3- Document Layout Analysis (Yolo V3-DLA) which is a Deep learning model to segment the various regions in the clinical document. (2) Identifying regions containing PHI through pattern-matching techniques and deleting or anonymizing the information in those regions. The proposed method was implemented to identify regions based on content structure, facilitating the de-identification of PHI regions and achieving an F1 score of 0.97. This system can be readily adapted to accommodate any form of clinical document.