Non-redundant association rules between diseases and medications: an automated method for knowledge base construction

BMC Medical Informatics and Decision Making

Research Article

Gabriel Nisand¹  Hassina Lefèvre¹  Nicolas Jay²  Erik A Sauleau³  Nicolas Meyer³  François Séverac³ 

[1] Groupe Méthode en Recherche Clinique, Service de Santé Publique, Hôpitaux Universitaires de Strasbourg, Strasbourg, France;LORIA – Equipe Orpailleurs, Université de Nancy, Nancy, France;SPI-EAO, Faculté de Médecine, Université de Nancy, Nancy, France;Laboratoire de Biostatistique et d’Informatique Médicale, Faculté de Médecine, Hôpitaux Universitaires de Strasbourg, Strasbourg, France;Groupe Méthode en Recherche Clinique, Service de Santé Publique, Hôpitaux Universitaires de Strasbourg, Strasbourg, France;

关键词: Data mining;    Association rules mining;    Natural language processing;    Knowledge base;   

DOI  :  10.1186/s12911-015-0151-9

 received in 2014-05-27, accepted in 2015-03-26,  发布年份 2015

来源: Springer

【 摘 要 】

BackgroundThe widespread use of electronic health records (EHRs) has generated massive clinical data storage. Association rules mining is a feasible technique to convert this large amount of data into usable knowledge for clinical decision making, research or billing. We present a data driven method to create a knowledge base linking medications to pathological conditions through their therapeutic indications from elements within the EHRs.MethodsAssociation rules were created from the data of patients hospitalised between May 2012 and May 2013 in the department of Cardiology at the University Hospital of Strasbourg. Medications were extracted from the medication list, and the pathological conditions were extracted from the discharge summaries using a natural language processing tool. Association rules were generated along with different interestingness measures: chi square, lift, conviction, dependency, novelty and satisfaction. All medication-disease pairs were compared to the Summary of Product Characteristics, which is the gold standard. A score based on the other interestingness measures was created to filter the best rules, and the indices were calculated for the different interestingness measures.ResultsAfter the evaluation against the gold standard, a list of accurate association rules was successfully retrieved. Dependency represents the best recall (0.76). Our score exhibited higher exactness (0.84) and precision (0.27) than all of the others interestingness measures. Further reductions in noise produced by this method must be performed to improve the classification precision.ConclusionsAssociation rules mining using the unstructured elements of the EHR is a feasible technique to identify clinically accurate associations between medications and pathological conditions.

【 授权许可】

CC BY   
© Séverac et al.; licensee BioMed Central. 2015

【 图 表 】

12864_2016_3440_Article_IEq67.gif