Medical multiple-choice question answering (MCQA) is a challenging evaluation for medical natural language processing and a helpful task in itself. Medical questions may describe patient symptoms and ask for the correct diagnosis, which requires domain knowledge and complex reasoning. Standard language modeling pretraining alone is not sufficient to achieve the best results with BERT-base size (Devlin et al., 2019) encoders. Jin et al. (2020) showed that focusing masked language modeling on disease name prediction when using medical encyclopedic paragraphs as input leads to considerable MCQA accuracy improvement. In this work, we show that (1) fine-tuning on generated MCQA dataset outperforms the masked language modeling based objective and (2) correctly masking the cues to the answers is critical for good performance. We release new pretraining datasets and achieve state-of-the-art results on 4 MCQA datasets, notably +5.7% with base-size model on MedQA-USMLE.

In recent years, the analysis of clinical texts has evolved significantly, driven by the emergence of language models like BERT such as PubMedBERT, and ClinicalBERT, which have been tailored for the (bio)medical domain that rely on extensive archives of medical documents. While they boast high accuracy, their lack of interpretability and language transfer limitations restrict their clinical utility. To address this, we propose a new, lightweight graph-based embedding method designed specifically for radiology reports. This approach considers the report’s structure and content, connecting medical terms through the multilingual SNOMED Clinical Terms knowledge base. The resulting graph embedding reveals intricate relationships among clinical terms, enhancing both clinician comprehension and clinical accuracy without the need for large pre-training datasets. Demonstrating the versatility of our method, we apply this embedding to two tasks: disease and image classification in X-ray reports. In disease classification, our model competes effectively with BERT-based approaches, yet it is significantly smaller and requires less training data. Additionally, in image classification, we illustrate the efficacy of the graph embedding by leveraging cross-modal knowledge transfer, highlighting its applicability across diverse languages.