MedNote BERT: Modeling Medical Notes and Predicting Drug-Disease Interactions

Electronic health records (EHRs) are critical repositories of patient health-related observations used in medicine, drug-drug interaction analysis, and medical research. However, the unstructured and unsystematic nature of EHR presents significant challenges in processing and analyzing this data. This paper introduces MedicineNote Biomedical Entity Representation Transformer (MedNote-BERT), a state-of-the-art model designed to overcome these challenges by integrating data from EHRs with the UMLS Metathesaurus. MedNote-BERT extracts textual embeddings from EHRs and employs graph learning model to construct a heterogeneous text graph, enriching the representation with biomedical concepts. The model further enhances learning by focusing on synonymous term pairs and employing hard sampling to refine its understanding. MedNote-BERT captures complex biomedical relationships, accurately predicting drug-drug side effects. The proposed methodology demonstrates significant improvements over existing benchmark models, highlighting MedNote-BERT's contribution to advancing the detection of adverse drug interactions from EHRs.