Method for Biomedical Information Extraction of Immunosuppressive Cell Properties

Automated extraction of cell populations' immunosuppressive properties from research articles is a basic problem, which requires specialized methods and tools for meta-analysis of publications. It is necessary to extract information about specific types of cells, their roles in the text, detect the immunosuppressive properties of cell populations, and certain types of relationships. It is also crucial to filter out those texts, which describe immunosuppressive features of different chemical compounds or drugs. Typically, efficient automatic information extraction requires a relatively large set of samples or marked-up texts. The paper presents a novel information extraction method that can be useful for such an analysis. This method uses external linguistic resources and can be trained on limited corpora. Namely, the method combines medical ontologies, rich unsupervised lexis representations (Fasttext word embeddings) with rule-based entity and relationship extraction, and supervised machine learning-based post-filtering. The developed method allows one to extract information about the target cells (for 'in vitro' experiments), effector cells, diseases ("in vivo'), and arbitrary descriptions of immune suppression. In the paper, we also present a manually labeled corpus to train immunosuppressive information extraction methods. That corpus contains texts of 330 PubMed Central abstracts. The experiments on that corpus show the method has relatively high evaluation scores on the labeled dataset. Therefore, the proposed method makes it possible to identify descriptions of the immunosuppressive properties of cell populations in biomedical texts with sufficiently high quality. In the future, the method can be applied to perform an automatic meta-analysis of research in immunosuppressive cell therapy. © 2021 IEEE.