Related Experiment Videos
A comparison of the abilities of nine scoring algorithms in predicting mortality
J Wayne Meredith1, Gregory Evans, Patrick D Kilgo
1Department of Surgery, Wake Forest University School of Medicine, Winston-Salem, North Carolina 27157, USA. merediw@wfubmc.edu
Objective:
The purpose of this study was to compare the abilities of nine Abbreviated Injury Scale (AIS)- and (ICD-9)-based scoring algorithms in predicting mortality.
Methods:
The scores collected on 76,871 incidents consist of four AIS-based algorithms (Injury Severity Score [ISS], New Injury Severity Score, Anatomic Profile Score [APS], and maximum AIS [maxAIS]), their four ICD to AIS mapped counterparts, and the ICD-9-based ISS (ICISS). A 10-fold cross-validation was performed and area under the receiver operating characteristic curve was used to determine algorithm discrimination. Hosmer-Lemeshow statistics were computed to gauge goodness-of-fit, and model refinement measured variance of predicted probabilities.
Results:
Overall, the ICISS has the best discrimination and model refinement, whereas the APS has the best Hosmer-Lemeshow performance. ICD-9 to AIS mapped scores have worse discrimination than their AIS-based counterparts, but still show moderate performance.
Conclusion:
Differences in performance were relatively small. Complex scores such as the ICISS and the APS provide improvement in discrimination relative to the maxAIS and the ISS. Trauma registries should move to include the ICISS and the APS. The ISS and maxAIS perform moderately well and have bedside benefits.