A multi-state model for evolution of intensive care unit patients: prediction of nosocomial infections and deaths

Nosocomial (hospital-acquired) infections are very frequent in intensive care units (ICU). The risk of death after severe infection is high, but the precise rate of death in ICU attributable to nosocomial infection is not known. The goal of this project was to build a statistical model to predict the occurrence of nosocomial infections in ICU and the outcome of the patients. We collected data on 676 consecutive patients admitted to an ICU for more than 24 hours between 1993 and 1996. The following data were collected for each patient: history; clinical examination at entry; subsequent infections; outcome. A multi-state heterogeneous semi-Markov model was determined and then validated; the initial data set was randomly split into two groups: two-thirds (450 patients) to build the model and one-third (226 patients) to validate it. The model defined five states: ICU admission; first simple infection; first complicated infection; death, and discharge from the ICU. Transitions between these states determined nine different events. The global model of patient histories can be divided into nine survival models, each corresponding to one of these events. The possible events from a given state were considered to be competing. Since many risk factors induced non-proportional hazard functions, piecewise exponential models were used to model event occurrence. The effect of continuous covariates on hazard functions has been described with a non-parametric method that enables non-linear relations to be shown. Among other things, the model allows patients' post-admission histories to be predicted from data available at ICU admission. The bootstrap estimator of the attributable risk of death due to simple or complicated nosocomial infections is 44.2 percent (95 percent CI 26.0-61.6 percent). We were also able to characterize the most highly exposed patients, those who comprise the high-risk group on whom prevention efforts must be focused.