Detecting related cases of bloodstream infections using time-interval distribution modelling.

An algorithm was designed to highlight related bloodstream infections using data from a nosocomial infection surveillance system to help local public health authorities direct specific measures towards clusters of cases. The approach was based on a two-step procedure. The first was a test to identify pathogens with an abnormal number of close cases. The second modelled, for the identified pathogens, the distribution of time intervals between successive cases as a mixture of two theoretical distributions in order to determine a threshold below which a specific investigation is required. The algorithm was applied to bloodstream infection surveillance data collected during a 10-year period (1996-2005) in an 878-bed teaching hospital (24 wards) in Lyon, France. The first step identified seven pathogens among the 18 being studied. The modelling succeeded in setting time thresholds to spot clusters of cases requiring further investigation with defined sensitivity and specificity. Setting the sensitivity level at 95%, the threshold values ranged from 24 days (Acinetobacter baumannii) to 294 days (Enterobacter cloacae); the specificity was higher than 70% (up to 97.5% for A. baumannii) except for E. cloacae (52.1%). Setting the specificity level at 95% resulted in a decrease in sensitivity except for A. baumannii (it reached nearly 100%); it fell below 50% for three pathogens: around 40% for Streptococcus pneumoniae and Enterococcus faecalis and 25% for Enterobacter cloacae. The threshold values then ranged from 8 days (S. pneumoniae) to 67 days (Streptococcus pyogenes). The approach proved promising though further refinements are needed before routine use.