Measure clinical drug-drug similarity using Electronic Medical Records.

OBJECTIVE: Quantitative measurement of clinical drug-drug similarity has many potential applications in assessing medication therapy similarity and patient similarity. Currently, most of the methods to measure drug-drug similarity were not directly obtained from clinical data and cannot cover clinical drugs. We sought to propose a computational approach to measure clinical drug-drug similarity based on the Electronic Medical Record (EMR) system.

MATERIALS AND METHODS: We used the Bonferroni-corrected hypergeometric P value to generate statistically significant associations between drugs and diagnoses in an EMR dataset which contained 812 554 medication records and 339 269 discharge diagnosis codes. Then the Jaccard similarity coefficient was used to measure the distances between drugs. A k-means based bootstrapping method was proposed to generate drug clusters.

RESULTS: The similarity matrix contains total 1210 clinical drugs used in the hospital was calculated. The clinical drug-drug similarity shows significant correlation with the chemical similarity of drugs and literature-based drug-drug similarity but with unique features. Based on this drug-drug similarity, 36 clinical drug clusters most of which were related to specific clinical conditions were generated. Detail of this drug clusters available at https://kb4md.org:4000/drugcluster.

DISCUSSION: This method provided a whole new view of the relationship among clinical drugs. Furthermore, it has the potential to evaluate the effectiveness of drug knowledge translation and provide quantitative knowledge resources for many applications such as treatment comparisons and patient similarity.

CONCLUSION: We proposed a clinical drug-drug similarity measurement that generated from clinical practice data and covers all clinical drugs.