Developing a Data-driven Medication Indication Knowledge Base using a Large Scale Medical Claims Database

Medication-indication knowledge base (KB) is useful for clinical care and also a key enabler for secondary use of observational health data. Over the years there are several indication KBs being developed, however, they were built based on curated data sources and thus may not reflect actual clinical practice. The longitudinal observational health data contain information about real world practice of medication indication, but were rarely used in KB construc- tion. A major challenge of leveraging them is the confounders in multi-medication multi-diagnoses relations. In this study, we proposed a sampling based approach that could explicitly handle the aforementioned confounders, and consequently detect more accurate medication-indication relations. Based on this method, we created a medication- indication KB that reflects actual clinical practice and has broad medication and indication coverages. Our work represents the first attempt to develop a medication-indication KB from a large scale observational health data in an automated and unsupervised manner