Extracting Disease-Symptom Relationships by Learning Syntactic Patterns from Dependency Graphs

摘要

Disease-symptom relationships are of primary importance for biomedical informatics, but databases that catalog them are incomplete in comparison with the state of the art available in the scientific literature. We propose in this paper a novel method for automatically extracting disease-symptom relationships from text, called SPARE (standing for Syntactic PAttern for Relationship Extraction). This method is composed of 3 successive steps: first, we learn patterns from the dependency graphs; second, we select best patterns based on their respective quality and specificity (their ability to identify only disease-symptom relationships); finally, the patterns are used on new texts for extracting disease-symptom relationships. We experimented SPARE on a corpus of 121,796 abstracts of PubMed related to 457 rare diseases. The quality of the extraction has been evaluated depending on the pattern quality and specificity. The best F-measure obtained is 55.65% (for specificity ≥ 0.5 and quality ≥ 0.5). To provide an insight on the novelty of disease-symptom relationship extracted, we compare our results to the content of phenotype databases (OrphaData and OMIM). Our results show the feasibility of automatically extracting disease-symptom relationships, including true relationships that were not already referenced in phenotype databases and may involve complex symptom descriptions.