Prediction of three outcome states from pediatric intensive care

To develop a method based on admission day data for predicting patient outcome status as independently functional, compromised functional, or dead.

Prospectively acquired development and validation samples.

A pediatric intensive care unit located in a tertiary care center.

Consecutive admissions (n = 1,663) for predictor development, and consecutive admissions (n = 1,153) for predictor validation.

Pediatric Risk of Mortality score, baseline Pediatric Overall Performance Category score, age, operative status, and primary diagnosis classified into ten organ systems and nine etiologies were recorded at the time of intensive care unit admission. Predictor was developed by stepwise polychotomous logistic regression analysis for the outcome functional, compromised, and dead. Model fit was evaluated by chi-square statistics; prediction performance was measured by the area under the receiver operating characteristic curve, and classification table analysis of observed vs. predicted outcomes.

The resulting predictor included Pediatric Risk of Mortality, baseline Pediatric Overall Performance Category, operative status, age, and diagnostic factors from four systems (cardiovascular, respiratory, neurologic, gastrointestinal), and six etiologies (infection, trauma, drug overdose, allergy/immunology, diabetes, miscellaneous/undetermined). Its application to the validation sample yielded good agreement between the total number expected and the observed outcomes for each state (chi-square = 3.16, 2 degrees of freedom, p = .206), with area indices of 0.96 +/- 0.01 for discrimination of fully functional vs. the combination of the two poor outcome states (compromised or death), and 0.94 +/- 0.02 for discrimination of fully or compromised functional vs. death. The 3 x 3 classification resulted in correct classification rates of 83.2%, 74.4%, and 81.3%, for the outcomes functional, compromised, and death, respectively.

Prediction of three outcome states using physiologic status, baseline functional level, and broad-based diagnostic groupings at admission is feasible and may improve the relevance of quality of care assessment.