Development of Pediatric Comorbidity Prediction Model

OBJECTIVE: To develop a comorbidity model for children that can be used with hospital discharge administrative databases. DESIGN: Retrospective study using administrative data obtained from the Canadian Institute for Health Information Discharge Abstract Database and the Deaths File to develop a logistic regression model. Hosmer-Lemeshow chi2 test was used to examine model fit. The C statistic was used to assess model discrimination. Bootstrapping was used to determine the stability of regression coefficients. SETTING: We used linked administrative databases to compile 339,077 hospital discharge abstracts from April 1, 1991, through March 31, 2002. PARTICIPANTS: Children between ages 1 and 14 years in Ontario, Canada. MAIN OUTCOME MEASURE: Death within 1 year of hospital discharge. RESULTS: The 27-variable pediatric comorbidity model predicted 1-year mortality with a C statistic of 0.83 in the Ontario data set from which it was derived. The presence of brain cancer (odds ratio, 76.38 [95% confidence interval, 53.40-109.27]) at hospital admission was the strongest predictor, followed by diabetes insipidus (odds ratio, 39.23 [95% confidence interval, 20.75-74.17]). CONCLUSION: Using clinical judgment and empirical modeling strategies, we were able to identify 27 diagnoses highly predictive of death for children between 1 and 14 years of age within 1 year of hospital discharge.