OBJECTIVES:: To build and validate a ventilator-associated pneumonia risk score for benchmarking. The rate of ventilator-associated pneumonia varies widely with case-mix, a fact that has limited its use for measuring intensive care unit performance.

METHODS:: We studied 1856 patients in the OUTCOMEREA database treated at intensive care unit admission by endotracheal intubation followed by mechanical ventilation for >48 hrs; they were allocated randomly to a training data set (n = 1233) or a validation data set (n = 623). Multivariate logistic regression was used. Calibration of the final model was assessed in both data sets, using the Hosmer-Lemeshow chi-square test and receiver operating characteristic curves.

MEASUREMENTS AND MAIN RESULTS:: Independent risk factors for ventilator-associated pneumonia were male gender (odds ratio = 1.97, 95% confidence interval = 1.32-2.95); SOFA at intensive care unit admission (<3 [reference value], 3-4 [2.57, 1.39-4.77], 5-8 [7.37, 4.24-12.81], >8 [5.81 (3.2-10.52)], no use within 48 hrs after intensive care unit admission of parenteral nutrition (2.29, 1.52-3.45), no broad-spectrum antimicrobials (2.11, 1.46-3.06); and mechanical ventilation duration (<5 days (); 5-7 days (17.55, 4.01-76.85); 7-15 days (53.01, 12.74-220.56); >15 days (225.6, 54.3-936.7). Tests in the training set showed good calibration and good discrimination (area under the curve-receiver operating characteristic curve = 0.881), and both criteria remained good in the validation set (area under the curve-receiver operating characteristic curve = 0.848) and good calibration (Hosmer-Lemeshow chi-square = 9.98, p = .5). Observed ventilator-associated pneumonia rates varied across intensive care units from 9.7 to 26.1 of 1000 mechanical ventilation days but the ratio of observed over theoretical ventilator-associated pneumonia rates was >1 in only two intensive care units.

CONCLUSIONS:: The ventilator-associated pneumonia rate may be useful for benchmarking provided the ratio of observed over theoretical rates is used. External validation of our prediction score is needed.