Prediction models for complications in trauma patients

Background: Because the death rate among the total trauma population is low, other performance indicators in addition to the classical dependent variable mortality are required to assess the overall quality of trauma care. The aim of this study was to develop and validate a prediction model for the occurrence of complications that can be used to adjust a measure of quality of trauma care for case mix. Methods: Complications recorded in a trauma registry between 1997 and 2008 were analysed. Formulas for different types of complication (institution- or diagnosis-related) derived from logistic regression models were used to calculate the probability of absence of complications (PAC). Discriminative power was tested by calculating the area under the receiver operating characteristic curve (AUC) in test and validation samples. Calibration was tested using Hosmer and Lemeshow methodology. Results: Some 5944 surgical trauma admissions were included in the analysis. A significant association between both institution- and diagnosis-related complications and Injury Severity Score was found. Diagnosis-related complications were also associated with Glasgow Coma Score and age. The AUCs of the PACs for institution- and diagnosis-related complications were 0.64 and 0.75 respectively in the test sample, and 0.66 and 0.76 in the validation sample. The AUCs increased when the outcomes of the models were divided into subcategories of complications. Hosmer and Lemeshow tests were not significant for all models, except that for institutional complications. Conclusion: To predict complications, a distinction should be made between institution- and diagnosis-related complications. The development of more detailed diagnosis-related prediction models is preferable because of better performance. The formulas predicting the PAC can be used to compare expected and observed complications.