Intelligent assessment method of VTE risk based on Lexicon-Biaffine named entity recognition

Venous thromboembolism (VTE) is characterized by high morbidity, mortality, and complex treatment, and its risk assessment is crucial in clinical care. Entities such as examinations, drugs, and treatments in electronic medical records provide important information for VTE risk assessment. However, there are a large number of specialized terms and abbreviations in Chinese EHRs, which can easily lead to the blurring of entity boundaries and the loss of semantic information. For this reason, this paper proposes a Named Entity Recognize (NER) method named Lexicon-Biaffine. The method combines medical dictionary matching technology and Biaffine Attention to accurately model word boundaries and their complex semantic relationships, and integrates BiLSTM and Temporal Convolutional Network (TCN) to model remote and local dependencies, respectively, so as to effectively mitigate the effect of semantic information loss. This method is experimentally demonstrated to achieve a significant improvement in F1 value, reaching 93.84%. In addition, based on the VTE risk factors extracted from the EHR, risk assessment of Padua, Caprini, Wells, and Geneva scales was realized using LLM reasoning, and the accuracy reached 88.32%, 79.56%, 90.51%, and 84.67%, respectively, based on the clinician's assessment results.