Experimental evaluation of bidirectional encoder representations from transformers models for de-identification of clinical document images

Abstract

Many health institutes maintain patients’ diagnosis and treatment reports as scanned images. For healthcare analytics and research, large volumes of digitally stored patient information have to be accessed, but the privacy requirements of protected health information (PHI) limit the research opportunities. Particularly in this artificial intelligence (AI) era, deep learning models require large datasets for training purposes, which hospitals cannot share unless the PHI fields are de-identified. Manual de-identification is beyond possible, with millions of patient records generated in hospitals every day. Hence, this work aims to automate the de-identification of clinical document images utilizing AI models, particularly pre-trained bidirectional encoder representations from transformers (BERT) models. For the purpose of experimentation, a synthetic dataset of 550 clinical document images was generated, encompassing data obtained from diverse patients across multiple hospitals. This work presents a two-stage transfer learning approach, initially employing Tesseract character recognition (OCR) to convert clinical document images into text. Subsequently, it extracts PHI fields from the text for de-identification. For the purpose of extraction, BERT models were utilized; in this work, we contrasted six pre-trained versions of such models to examine their effectiveness and achieve the F1 score of 92.45%, thus showing better potential for de-identifying PHI data in clinical documents.