Data Mining in MEDLINE for Disease-Disease Associations Via Second Order Co-Occurrence

DDMiner, a new method for mining disease-disease associations in MEDLINE, is presented together with its first results. DDMiner searches for co-occurrences of gene names and disease terms, and finds relationships between diseases by word vector-similarity calculations. All records in PubMed were labeled with around 40,000 gene and protein names, and around 4,000 disease terms. Each disease term was described by a word vector from which the length equals the number of gene names. Each field in the vector represented a gene or a protein. The value in the field was derived from the number of publications in which this gene occurred together with the disease term. Disease-disease associations were calculated by vector-similarity calculation. Five diseases were examined together with their closest neighbor diseases to show the validity of our approach. All five examples showed only disease-disease associations that could be validated by medical literature. These results show that mining for disease-disease associations by second order co-occurrence is a powerful tool for medical science.