Abstract
OBJECTIVE: Mining disease-specific associations from existing knowledge resources can be useful for building disease-specific ontologies and supporting knowledge-based applications. Many association mining techniques have been exploited. However, the challenge remains when those extracted associations contained much noise. It is unreliable to determine the relevance of the association by simply setting up arbitrary cut-off points on multiple scores of relevance; and it would be expensive to ask human experts to manually review a large number of associations. We propose that machine-learning-based classification can be used to separate the signal from the noise, and to provide a feasible approach to create and maintain disease-specific vocabularies. METHOD: We initially focused on disease-medication associations for the purpose of simplicity. For a disease of interest, we extracted potentially treatment-related drug concepts from biomedical literature citations and from a local clinical data repository. Each concept was associated with multiple measures of relevance (i.e., features) such as frequency of occurrence. For the machine purpose of learning, we formed nine datasets for three diseases with each disease having two single-source datasets and one from the combination of previous two datasets. All the datasets were labeled using existing reference standards. Thereafter, we conducted two experiments: (1) to test if adding features from the clinical data repository would improve the performance of classification achieved using features from the biomedical literature only, and (2) to determine if classifier(s) trained with known medication-disease data sets would be generalizable to new disease(s). RESULTS: Simple logistic regression and LogitBoost were two classifiers identified as the preferred models separately for the biomedical-literature datasets and combined datasets. The performance of the classification using combined features provided significant improvement beyond that using biomedical-literature features alone (p-value<0.001). The performance of the classifier built from known diseases to predict associated concepts for new diseases showed no significant difference from the performance of the classifier built and tested using the new disease's dataset. CONCLUSION: It is feasible to use classification approaches to automatically predict the relevance of a concept to a disease of interest. It is useful to combine features from disparate sources for the task of classification. Classifiers built from known diseases were generalizable to new diseases.