ICISS: an international classification of disease-9 based injury severity score

The Injury Severity Score (ISS) has served as the standard summary measure of human trauma for 20 years. Despite its stalwart service, the ISS has two weaknesses: it relies upon the consensus derived severity estimates for each Abbreviated Injury Scale (AIS) injury and considers, at most, only three of an individual patient's injuries, three injuries that often are not even the patient's most severe injuries. Additionally, the ISS requires that all patients have their injuries described in the AIS lexicon, an expensive step that is currently taken only at hospitals with a zealous commitment to trauma care. We hypothesized that a data driven alternative to ISS that used empirically derived injury severities and considered all of an individual patient's injuries would more accurately predict survival.

Survival risk ratios were derived for every International Classification of Disease 9th Edition (ICD-9) injury category (800-959.9) using the North Carolina State Discharge Database experience with 300,000 trauma patients over 5 years. An ICD-9 Injury Severity Score (ICISS) was then defined as the product of all survival risk ratios for an individual patient's traumatic ICD-9 codes. We compared the performance of ISS and ICISS in a group of 3,142 patients accrued at the University of New Mexico Trauma Center over 4 years. These patients had both AIS and ICD-9 descriptors meticulously assigned prospectively by designated trauma data base personnel.

ICISS outperformed ISS at a level that was highly statistically significant (p < 0.0001) and may be clinically important: ISS misclassification rate 7.67%, ISS Receiver Operator Characteristic Curve area = 0.872; ICISS misclassification rate 5.95%, ICISS Receiver Operator Characteristic Curve area = 0.921. Moreover, these improvements are largely preserved when ICISS is used in a probability of survival model that includes age, mechanism, and revised trauma score. About half of ICISS's improvement in predictive power is because of its use of an individual patient's worst three injuries regardless of body region. The remainder is because of better modeling of individual injuries and allowing all injuries to contribute to the final score.

We conclude that ICISS is a much better predictor of survival than ISS in injured patients. The use of the ICD-9 lexicon may avoid the need for AIS coding, and thus may add an economic incentive to the statistical appeal of ICISS. It is possible that a similar data driven revision of ISS using the AIS vocabulary might perform as well or better than ICISS. Indeed, the actual lexicon used to divide up the injury "landscape" into individual injuries may be of little consequence so long as all injuries are considered and empirically derived SRRs are used to calculate the final injury measure.