Outcome of older patients with severe pneumonia predicted by recursive partitioning

OBJECTIVES: To develop a prognostic model to predict outcome of older patients with severe pneumonia requiring mechanical ventilation. DESIGN: A nonconcurrent prospective study. SETTING: A 24-bed intensive care unit (ICU) within two university-affiliated tertiary care hospitals. PARTICIPANTS: All patients age 75 and older with severe pneumonia between June 1996 and September 1999 were included. Demographic data including activities of daily living (ADL) index score before admission, and clinical and laboratory data were collected in the first 24 hours of admission to the ICU. One hundred four patients (mean age ± standard deviation (SD) 82.3 ± 5.5 years) met the inclusion criteria. MEASUREMENTS: A classification tree was developed using binary recursive partitioning to predict hospital discharge. The model was compared with a logistic regression model using variables selected by the tree analysis and with the Acute Physiologic and Chronic Health Evaluation (APACHE) II. RESULTS: Outcome predictors for the classification tree were use of vasopressors, presence of multilobar pneumonia on chest radiograph, ratio of blood urea nitrogen to creatinine, Glasgow Coma Scale, urine output, and ADL score before admission. The tree achieved a sensitivity of 83.8% (95% confidence interval (CI) 69.2-92.4) and a specificity of 93.3% (95% CI 83-98.1). The predictive accuracy as assessed by the area under the curve (c-index ± standard error) was significantly higher with the classification tree (0.932 ± 0.03) than with logistic regression and APACHE II, (0.801 ± 0.028 and 0.711 ± 0.049, respectively (P < .05). CONCLUSIONS: The classification tree model demonstrated a superior predictive accuracy to that of logistic regression and APACHE II. If validated prospectively, the classification tree can be used as a tool to assess the outcome of older patients with severe pneumonia requiring mechanical ventilation on admission to the ICU. In addition, the classification tree can be used to assist healthcare workers in providing a concise summary of local outcome experience and prognostic information to patients and their surrogates.