Predictive model for nosocomial pneumonia in intensive care units

BACKGROUND AND OBJECTIVE: Nosocomial pneumonia is the most common nosocomial infection in the intensive care units (ICUs) and contributes disproportionately to both poor outcomes and high cost of care in critically ill patients. In order to identify patients with greater risk of developing nosocomial pneumonia in ICUs, it is important to select the right preventive measures. PATIENTS AND METHOD: It was an observational study of 2 prospective cohorts of patients staying in the ICU for 24 h or more: the main cohort (n = 1,184) and the validation cohort (n = 554). A predictive model was constructed with the data of the main cohort using a logistic regression. Receiver operating characteristic (ROC) curves and predictive values for different cut points were obtained with the data of both cohorts. RESULTS: Eight variables were selected for the predictive model: parenteral nutrition, enteral nutrition, nasogastric intubation, tracheostomy, mechanical ventilation, previous surgery, coma and diabetes. In the main cohort, the model had a sensitivity of 81% and a specificity of 78.4% in predicting nosocomial pneumonia (Hosmer-Lemeshow statistic p = 0.93; area under ROC curve = 0.861; 95% confidence interval, 0.824-0.898). In the validation cohort, the area under ROC curve was 0.849 (95% confidence interval, 0.742-0.956). CONCLUSIONS: The resulting model presents satisfactory results in both cohorts. In addition, the variables used are simple, routinely available, and familiar to clinicians.