Nosocomial infections: validation of surveillance and computer modeling to identify patients at risk

Abstract

To estimate the accuracy of routine hospital-wide surveillance for nosocomial infection, the authors performed a validation study at the University of Iowa Hospitals and Clinics, a 900-bed tertiary care institution, by daily concurrent surveys of all patients' charts. The study extended over a 10-month period from January to October 1987. The sensitivity and specificity of the reported data were 80.7% (95% confidence interval (CI) 72.2-89.2) and 97.5% (95% CI 96.4-98.5), respectively. The predictive values of positive or negative reports of an infection were 75.3% (95% CI 66.3-84.2) and 98.1% (95% CI 97.3-99.1), respectively. In a separate analysis, the data entry system was reviewed for eight descriptive variables among all patients with infections (n = 443) identified over a 2-month period. The data entry was found to be 94-99% accurate. To improve the efficiency of current surveillance, the authors used data gathered during the study to develop a computer model for the identification of patients with a high probability of having a nosocomial infection. The use of stepwise logistic regression identified five variables which independently predicted infection: age of the patient (years), days of antibiotics, days of hospitalization, and the number of days on which urine and/or wound cultures were obtained. Optimal sensitivity and specificity (81.6% and 72.5%, respectively) were found when the model examined patients with an 8% or higher a priori probability of infection; this figure corresponded to a review of 33% of the patients' charts. Increasing the a priori probability would progressively increase specificity and reduce both sensitivity and the number of charts needed for review. If it is prospectively validated, the model may provide a more efficient mechanism by which to conduct hospital-wide surveillance.