Customized Impression Prediction From Radiology Reports Using BERT and LSTMs
Batuhan Gündoğdu, Utku Pamuksuz, Jonathan H. Chung, Jessica M. Telleria, Peng Liu, Farrukh Aslam Khan, Paul J. Chang
Abstract
Clinical language processing has become an attractive field with the improvements of deep learning applications and the abundance of large unstructured narratives in the healthcare records. The capability to extract unstructured information from raw text to provide actionable information for healthcare personnel plays a vital role in healthcare workflows. In this article, we introduce a deep learning approach to automate the generation of radiology impressions by analyzing radiology findings and patient background information of each examination. Since the impression section of a radiology report is an essential conclusion, any errors can prove to be detrimental. Thus, we developed a deep learning system to prevent important clinical findings from being overlooked by using almost 1 million de-identified radiology reports obtained from the University of Chicago Medicine over the last 12 years. We propose to automate the generation of radiology reports by incorporating sequence-to-sequence neural network models with the power of bidirectional encoder representations from transformers (BERT). We tested our model in a real-time experimental setup with radiologists in a top tier academic institution and statistically validated the performance by using ROUGE metrics. Clinical validations have shown that 76% of our predictions are at least as accurate as human-generated impressions by radiologists. Furthermore, statistical validation metrics demonstrated higher ROUGE scores compared to previously published studies over two different test sets.