LAUSR

BACKGROUND Hospital benchmarking is essential to quality improvement, but its usefulness depends on the ability of statistical models to adequately control for inter-hospital differences in patient mix. We explored whether the addition of injury-specific clinical variables to the current American College of Surgeons-Trauma Quality Improvement Program (TQIP) algorithm would improve model fit. METHODS We analyzed a prospective registry containing all adult patients who presented to a regional consortium of 14 trauma centers between 2010-2011 with severe traumatic brain injury (TBI). We used hierarchical logistic regression and stepwise forward selection to develop two novel risk-adjustment models. We then tested our novel models against the current TQIP model and ranked hospitals by their risk-adjusted mortality rates under each model to determine how model selection affects quality benchmarking. RESULTS 734 patients met inclusion criteria. Stepwise selection resulted in two distinct models: one that added three TBI-specific variables (pupil reactivity, cerebral edema, loss of basal cisterns) to the model specification currently used by TQIP and another that combined two TBI-specific variables (pupil reactivity, cerebral edema) with a three-variable subset of TQIP (age, Abbreviated Injury Scale score for the head region, Glasgow Coma Scale motor score). Both novel models outperformed TQIP. Although rankings remained largely unchanged across model configurations, several hospitals moved across quality terciles. CONCLUSION The inclusion of injury-specific variables improves risk adjustment for patients with severe TBI. TQIP should consider replacing several of its general patient characteristics with injury-specific clinical predictors to increase efficiency, reduce the risk of overfitting, and improve the accuracy of hospital benchmarking. LEVEL OF EVIDENCE Prognostic and Epidemiological, Level II.