Discovering Outliers of Potential Drug Toxicities Using a Large-scale Data-driven Approach

Jake Luo; Ron A Cisler

doi:10.4137/CIN.S39549

Discovering Outliers of Potential Drug Toxicities Using a Large-scale Data-driven Approach

Cancer Inform. 2016 Oct 26:15:211-217. doi: 10.4137/CIN.S39549. eCollection 2016.

Authors

Jake Luo¹, Ron A Cisler²

Affiliations

¹ Center for Biomedical Data and Language Processing, Department of Health Informatics and Administration, University of Wisconsin-Milwaukee, Milwaukee, WI, USA.; Department of Health Informatics and Administration, College of Health Sciences, University of Wisconsin-Milwaukee, Milwaukee, WI, USA.; College of Health Sciences, University of Wisconsin-Milwaukee, Milwaukee, WI, USA.
² Center for Biomedical Data and Language Processing, Department of Health Informatics and Administration, University of Wisconsin-Milwaukee, Milwaukee, WI, USA.; Department of Health Informatics and Administration, College of Health Sciences, University of Wisconsin-Milwaukee, Milwaukee, WI, USA.; College of Health Sciences, University of Wisconsin-Milwaukee, Milwaukee, WI, USA.; Center for Urban Population Health, Milwaukee, WI, USA.; Joseph J. Zilber School of Public Health, University of Wisconsin-Milwaukee, Milwaukee, WI, USA.

Abstract

We systematically compared the adverse effects of cancer drugs to detect event outliers across different clinical trials using a data-driven approach. Because many cancer drugs are toxic to patients, better understanding of adverse events of cancer drugs is critical for developing therapies that could minimize the toxic effects. However, due to the large variabilities of adverse events across different cancer drugs, methods to efficiently compare adverse effects across different cancer drugs are lacking. To address this challenge, we present an exploration study that integrates multiple adverse event reports from clinical trials in order to systematically compare adverse events across different cancer drugs. To demonstrate our methods, we first collected data on 186,339 clinical trials from ClinicalTrials.gov and selected 30 common cancer drugs. We identified 1602 cancer trials that studied the selected cancer drugs. Our methods effectively extracted 12,922 distinct adverse events from the clinical trial reports. Using the extracted data, we ranked all 12,922 adverse events based on their prevalence in the clinical trials, such as nausea 82%, fatigue 77%, and vomiting 75.97%. To detect the significant drug outliers that could have a statistically high possibility of causing an event, we used the boxplot method to visualize adverse event outliers across different drugs and applied Grubbs' test to evaluate the significance. Analyses showed that by systematically integrating cross-trial data from multiple clinical trial reports, adverse event outliers associated with cancer drugs can be detected. The method was demonstrated by detecting the following four statistically significant adverse event cases: the association of the drug axitinib with hypertension (Grubbs' test, P < 0.001), the association of the drug imatinib with muscle spasm (P < 0.001), the association of the drug vorinostat with deep vein thrombosis (P < 0.001), and the association of the drug afatinib with paronychia (P < 0.01).

Keywords: adverse events; big clinical data; cancer informatics; clinical trial; drug toxicity; outlier discovery.