Attributable mortality of ventilator-associated pneumonia: Respective impact of main characteristics at ICU admission and VAP onset using conditional logistic regression and multi-state models

Abstract

Purpose: Methods for estimating the excess mortality attributable to ventilator-associated pneumonia (VAP) should handle VAP as a time-dependent covariate, since the probability of experiencing VAP increases with the time on mechanical ventilation. VAP-attributable mortality (VAP-AM) varies with definitions, case-mix, causative microorganisms, and treatment adequacy. Our objectives here were to compare VAP-AM estimates obtained using a traditional cohort analysis, a multistate progressive disability model, and a matched-cohort analysis; and to compare VAP-AM estimates according to VAP characteristics. Methods: We used data from 2,873 mechanically ventilated patients in the Outcomerea® database. Among these patients from 12 intensive care units, 434 (15.1%) experienced VAP; of the remaining patients, 1,969 (68.5%) were discharged alive and 470 (16.4%) died. With the multistate model, VAP-AM was 8.1% (95% confidence interval [95%CI], 3.1-13.1%) for 120 days' complete observation, compared to 10.4% (5.6-24.5%) using a matched-cohort approach (2,769 patients) with matching on mechanical ventilation duration followed by conditional logistic regression. VAP-AM was higher in surgical patients and patients with intermediate (but not high) Simplified Acute Physiologic Score II values at ICU admission. VAP-AM was significantly influenced by time to VAP but not by resistance of causative microorganisms. Higher Logistic Organ Dysfunction score at VAP onset dramatically increased VAP-AM (to 31.9% in patients with scores above 7). Conclusion: A multistate model that appropriately handled VAP as a time-dependent event produced lower VAP-AM values than conditional logistic regression. VAP-AM varied widely with case-mix. Disease severity at VAP onset markedly influenced VAP-AM; this may contribute to the variability of previous estimates. © 2010 Copyright jointly held by Springer and ESICM.