Development of statistical methodologies and risk models to perform real-time safety monitoring in interventional cardiology

Abstract

Post-marketing surveillance of medical pharmaceuticals and devices has received a great deal of media, legislative, and academic attention in the last decade. Among medical devices, these have largely been due to a small number of highly publicized adverse events, some of them in the domain of cardiac surgery and interventional cardiology. Phase three clinical trials for these devices are generally underpowered to detect rare adverse event rates, are performed in near-optimal environments, and regulators face significant pressure to deliver important medical devices to the public in a timely fashion. All of these factors emphasize the importance of systematic monitoring of these devices after being released to the public, and the FDA and other regulatory agencies continue to struggle to perform this duty using a variety of voluntary and mandatory adverse event rate reporting policies. Data quality and comprehensiveness have generally suffered in this environment, and delayed awareness of potential problems. However, a number of mandatory reporting policies combined with improved standardization of data collection and definitions in the field of interventional cardiology and other clinical domains have provided recent opportunities for nearly "real-time" safety monitoring of medical device data.

(cont.) Existing safety monitoring methodologies are non-medical in nature, and not well adapted to the relatively heterogeneous and noisy data common in medical applications. A web-based database-driven computer application was designed, and a number of experimental statistical methodologies were adapted from non-medical monitoring techniques as a proof of concept for the utility of an automated safety monitoring application. This application was successfully evaluated by comparing a local institution's drug-eluting stent in-hospital mortality rates to University of Michigan's bare-metal stent event rates. Sensitivity analyses of the experimental methodologies were performed, and a number of notable performance parameters were discovered. In addition, an evaluation of a number of well-validated external logistic regression models, and found that while population level estimation was well-preserved, individual estimation was compromised by application to external data. Subsequently, exploration of an alternative modeling technique, support vector machines, was performed in an effort to find a method with superior calibration performance for use in the safety monitoring application.