Validation of four burn-specific prognostic models in a cohort of 9625 cases, and a novel model for prediction of mortality in burn patients

Abstract

BackgroundSurvival after burn injury has steadily improved in recent decades. The models for assessing the severity of burn injury and predicting burn-associated mortality have been used for over 20 years. The predictive accuracy of these models should be reconsidered now. MethodIn this retrospective study on all burn patients (n = 9625) admitted to the Burn Department, Southwest Hospital between 2008 and 2017, we compared the predictive performance of the four burn-severity models (Abbreviated Burn Severity Index, Ryan score, revised Baux score and Belgian Outcome of Burn Injury) by area under the receiver operating curve (AUC) and Hosmer–Lemeshow test. We developed a new model with the data from 2008 to 2012 (5006 patients) by logistic regression, data from 2013 to 2017 (4619 patients) were used for validation. ResultThe overall mortality rate of the burn patients was 1.14%. The four previously validated burn models showed good discrimination power of death risk (AUC > 0.890) but poor fitness to the observed mortality rate (p < 0.001). Risk factors associated with mortality included sex, age, total burn area, full thickness burn area, and inhalation injury. The new logistic model was devised with high sensitivity and specificity (0.913 and 0.806, respectively) and an AUC of 0.940. The new model also had good fitness to the observed mortality of burn patients (p = 0.588). ConclusionThe four widely used burn models have poor accuracy in predicting burn-associated mortality, and an accurate new model was developed based on simple and objective clinical characteristics of burn patients at admission.