A BART-Based Study of Entity-Relationship Extraction for Electronic Medical Records of Cardiovascular Diseases

With the advancement of training techniques, models such as BERT and GPT have been pre-trained on massive unlabeled texts, enabling effective learning of semantic representations of sentences and incorporating a wealth of prior knowledge. Therefore, we utilized a pre-trained model based on BART, called Joint entity Relation Extraction with BART (JREwBART), for entity relation extraction in medical texts. Building upon the JREwBART model, we proposed the Pipeline entity Relation Extraction based on the BART and Biaffine Transformation (PRE-BARTaBT) model. We evaluated the performance of these two models on three Chinese medical datasets: SEMRC, CVDEMRC (Cardiovascular/Stroke Disease Electronic Medical Record entity and relation Corpus), and CMeIE. The experimental results demonstrate the effectiveness of both models. Compared to the state-of-the-art baseline model, Cas-CLN, JREwBART achieved an improvement of \(0.71\%\) , \(1.64\%\) , and \(0.37\%\) in terms of F1 score on the three datasets, respectively. PRE-BARTaBT showed F1 score improvements of \(0.81\%\) , \( 2\%\) , and \(0.26\%\) on the same datasets, respectively.