An Empirical Evaluation of Resources for the Identification of Diseases and Adverse Effects in Biomedical Literature

The mentions of human health perturbations such as the diseases and adverse effects denote a special entity class in the biomedical literature. They help in understanding the underlying risk factors and develop a preventive rationale. The recognition of these named entities in texts through dictionary-based approaches relies on the availability of appropriate terminological resources. Although few resources are publicly available, not all are suitable for the text mining needs. Therefore, this work provides an overview of the well known resources with respect to human diseases and adverse effects such as the MeSH, MedDRA, ICD-10, SNOMED CT, and UMLS. Individual dictionaries are generated from these resources and their performance in recognizing the named entities is evaluated over a manually annotated corpus. In addition, the steps for curating the dictionaries, rule-based acronym disambiguation and their impact on the dictionary performance is discussed. The results show that the MedDRA and UMLS achieve the best recall. Besides this, MedDRA provides an additional benefit of achieving a higher precision. The combination of search results of all the dictionaries achieve a considerably high recall. The corpus is available on http://www.scai.fraunhofer.de/disease-ae-corpus.html