Joint Mixed Membership Clustering for Identifying Adverse Drug Events
COM1 Level 3
MR1, COM1-03-19
Examiners: Professor Atreyi Kankanhalli and Assistant Professor Stanley Kok
Abstract:
Adverse Drug Events (ADEs) are unintended drug effects that often lead to emergency visits, prolonged hospital stays and worse patient outcomes. They result in significant clinical and economic burden, that can be reduced if unknown ADEs are discovered early. Clinical Trials do not detect all ADEs since they do not evaluate effects of drugs on large cohorts, over longer periods of time and over multiple populations. So, post-marketing drug safety surveillance through voluntary reports and data mining of clinical databases is routinely carried out. Most previous studies have investigated ADEs of specific drugs or pairs of drugs. ADEs can be caused due to multiple combinations of drugs, but it is computationally intractable to test for ADEs of all combinations of a large set of drugs. Furthermore, ADEs have complex manifestations and not all ADEs are observed in all patient groups. Our goal is to discover groups of patients with ADEs to a combination of drugs by mining Electronic Health Records (EHR), to identify both new ADEs as well as the clinical characteristics of the patients who are afflicted by these ADEs. We design a framework using mixed membership clustering to automatically discover both groups of drugs and groups of diseases in EHR data and associate them through patient subgroups in a data-driven and computationally tractable manner. Our preliminary results are promising - the clusters discovered by our framework find potential ADEs of multiple combinations of drugs. For these drugs, the precision of our approach is found to be superior to methods commonly used for ADE detection.