Statistical Methods in Monitoring Adverse Events for Medical Device Postmarketing Surveillance Studies*

In 1984, the United States Food and Drug Administration (FDA) required manufacturers of medical and radiation-emitting devices to submit reports of events that reasonably suggest that a medical device has caused or may have caused or contributed to the death, serious injury, or serious illness of a patient, as well as malfunctions whose recurrence would likely result in death, serious injury, or serious illness of a patient. Like adverse drug reactions (ADRs), the medical device reports (MDRs) usually do not have the device usage (denominator) data available for monitoring and statistical analyses. MDRs from various sources (manufacturers, physicians, public health agencies, and others), reported to the FDA's Center for Devices and Radiological Health (CDRH), are regularly reviewed and processed. In this paper, we analyze the monthly MDRs during January 1990 to December 1993 (48 months). To facilitate the statistical analyses, we assume that the device usage data (denominator) are relatively constant during the study period. In this paper, our interest includes MDR trend analysis, sudden spike signal detection or early warning, and the upper threshold cutoff point determination for MDR monitoring. Statistical methods used for the above issues include nonparametric regressions (polynomial regression fitting, lowess robust locally fit, kernel smoothing, and cubic spline smoothing) for exploratory trend analysis, neighborhood MDR pairwise comparison for sudden spike detection, Cox-Stuart nonparametric trend test, and various statistical models (mixed Poisson/binomial, zero truncated Poisson, and negative binomial or compound Poisson) for MDR monitoring and upper threshold cutoff point determination. Due to non-stationary time series, excessive zeros or ones, and over-dispersion of the observed MDRs, the time series analysis, regression analysis, simple Poisson, and cumulative sum (CUSUM) procedures are usually not appropriate. The two data sets used to illustrate the above statistical analyses consist of the observed 48-month reported MDR data from intravenous tube (IV) total (sum of death, injury or illness, and device malfunction) and hip implant injury reports from January 1990 through December 1993.