Scoping review of knowledge graph applications in biomedical and healthcare sciences

Abstract

AbstractIntroductionThere is increasing use of knowledge graphs within medicine and healthcare, but a comprehensive survey of their applications in biomedical and healthcare sciences is lacking. Our primary aim is to systematically describe knowledge graph use cases, data characteristics, and research attributes in the academic literature. Our secondary objective is to assess the extent of real-world validation of findings from knowledge graph analysis.MethodsWe conducted this review in accordance with the PRISMA extension for Scoping Reviews to characterize biomedical and healthcare uses of knowledge graphs. Using keyword-based searches, relevant publications and preprints were identified from MEDLINE, EMBASE, medRxiv, arXiv, and bioRxiv databases. A final set of 255 articles were included in the analysis.ResultsAlthough medical science insights and drug repurposing are the most common uses, there is a broad range of knowledge graph use cases. General graphs are more common than graphs specific to disease areas. Knowledge graphs are heterogenous in size with median node numbers 46 983 (IQR 6 415-460 948) and median edge numbers 906 737 (IQR 66 272-9 894 909). DrugBank is the most frequently used data source, cited in 46 manuscripts. Analysing node and edge classes within the graphs suggests delineation into two broad groups: biomedical and clinical. Querying is the most common analytic technique in the literature; however, more advanced machine learning techniques are often used.DiscussionThe variation in use case and disease area focus identifies areas of opportunity for knowledge graphs. There is diversity of graph construction and validation methods. Translation of knowledge graphs into clinical practice remains a challenge. Critically assessing the success of deploying insights derived from graphs will help determine the best practice in this area.