Transfer learning with BERT and a-priori Knowledge-Based Sentence of Interest Selection in Radiology Impressions for Phenotyping Venous Thromboembolism

Abstract

Phenotyping venous thromboembolism (VTE) is a challenging task that requires accurate identification of clinical features from unstructured electronic health records (EHRs). In this study, we propose the use of Bidirectional Encoder Representations from Transformers (BERT), a pre-trained natural language processing (NLP) model, for VTE phenotyping. We fine-tuned BERT on a corpus consisting of radiology impressions of 13702 cancer patients from Harris Health System (HHS) in Houston, Texas. Our evaluation shows that BERT can achieve a sensitivity of 96.1% and precision of 95.1%. Our findings indicate that BERT can be an effective tool for VTE phenotyping using radiology impressions. The proposed approach has potential applications in clinical decision support and population health management.