KeyCARE: a framework for biomedical Keyword Extraction, term Categorization, and semantic Relation

Abstract

The medical sector generates vast amounts of unstructured data, which, if processed correctly, can significantly enhance medical processes and their outcomes. This thesis presents the development of KeyCARE, a Python library for keyword extraction, term categorization, and relations extraction that tackles this need. Utilizing mainly unsupervised and few-shot methods, KeyCARE efficiently extracts classified keywords from medical records with a recall of up to 98% and an f-score of up to 61%, with partial overlaps considered as correct. While these scores are not comparable to those of supervised Named Entity Recognition systems, they set a high standard for an unsupervised alternative in scenarios of data scarcity. Moreover, the library incorporates relation extractors that identify hierarchical relationships among biomedical keywords and with terminologies, achieving a precision and recall of 93%. This has a clear application in terminology enrichment, data generation and information extraction, particularly in specific domains and low-resource languages such as Catalan. This thesis encompasses the comprehensive development of KeyCARE, including an in-depth evaluation of the implemented models as well as basic use cases demonstrating its practical applications.