Comparison of Trauma Mortality Prediction Models With Updated Survival Risk Ratios in Korea.

Abstract

Background: Despite the considerable disease burden due to trauma injury, sufficient effort has not been made for the assessment of nationwide trauma care status in Korea. We explored the feasibility of a diagnosis code-based injury severity measuring method in light of its real-world usage.

Methods: We used datasets from the National Emergency Department Information System to calculate the survival risk ratios (SRRs) and the Korean Trauma Data Bank to predict models, respectively. The target cohort was split into training and validation datasets using stratified random sampling in an 8:2 ratio. We established six major mortality prediction models depending on the included parameters: 1) the Trauma and Injury Severity Score (TRISS) (age, sex, original Revised Trauma Score [RTS], Injury Severity Score [ISS]), 2) extended International Classification of Diseases-based Injury Severity Score (ICISS) 1 (age, sex, original RTS, ICISS using international SRRs), 3) extended ICISS 2 (age, sex, original RTS, ICISS using Korean SRRs based on 4-digit diagnosis codes), 4) extended ICISS 3 (age, sex, original RTS, ICISS using Korean SRRs based on full-digit diagnosis codes), 5) extended ICISS 4 (age, sex, modified RTS, and ICISS using Korean SRRs based on 4-digit diagnosis codes), 6) extended ICISS 5 (age, sex, modified RTS, and ICISS using Korean SRRs based on full-digit diagnosis codes). We estimated the model using training datasets and fitted it to the validation datasets. We measured the area under the receiver operating characteristic curve (AUC) for discriminative ability. Overall performance was also evaluated using the Brier score.

Results: We observed the feasibility of the extended ICISS models, though their performance was slightly lower than the TRISS model (training cohort, AUC 0.936-0.938 vs. 0.949). Regarding SRR calculation methods, we did not find statistically significant differences. The alternative use of the Alert, Voice, Pain, Unresponsive Scale instead of the Glasgow Coma Scale in the RTS calculation did not degrade model performance.

Conclusion: The availability of the practical ICISS model was observed based on the model performance. We expect our ICISS model to contribute to strengthening the Korean Trauma Care System by utilizing mortality prediction and severity classification.