The comparison of injury severity instrument performance using likelihood ratio and ROC curve analyses.

Abstract

The relative merits of injury severity instruments are often difficult to determine because the statistical techniques, study populations, and outcomes used in trauma research tend to obscure rather than elucidate differences in performance. To describe the advantages of likelihood ratio and receiver operator characteristic (ROC) curve analyses and to demonstrate them using study populations and outcomes that facilitate instrument discrimination. Previously published data on the performance of the "Injury Severity Score" and of four triage instruments in the prediction of mortality; paired comparison of two instruments in a previously unpublished trauma registry dataset; use of likelihood ratio and ROC analyses. In a comparison of triage instruments this study clearly showed that CRAMS, PHI, and RTI, which contain nonphysiological information, have higher performance levels than the Revised Trauma Score (RTS), which is restricted to physiological information. Absolute performance gains of these instruments over the RTS ranged from 5.9% to 43.5%. This study also noted that the ISS performs equally well for blunt and penetrating trauma, and affirms concerns about the adequacy of its performance. Likelihood ratio and ROC curve analyses demonstrate differences in injury severity instrument performance that were obscured by less rigorous methods of comparison.