Related Experiment Videos
A simple mathematical modification of TRISS markedly improves calibration
Turner M Osler1, Frederick B Rogers, Gary J Badger
1University of Vermont, Fletcher, Allen Health Care, Burlington 05401, USA. turner.osler@vtmednet.org
The Journal of Trauma
|October 24, 2002
Summary
The Trauma and Injury Severity Score (TRISS) model has poor calibration due to the nonlinear relationship between Injury Severity Score (ISS) and mortality. Adding a quadratic ISS term significantly improves TRISS model calibration for trauma outcome prediction.
Area of Science:
- Trauma outcome prediction
- Biostatistics
- Medical informatics
Background:
- The TRISS model has been the standard for trauma outcome prediction for two decades.
- Despite its widespread use, TRISS exhibits poor calibration across various datasets.
- This study investigates the mathematical specification of the TRISS model as a potential cause for its calibration issues.
Purpose of the Study:
- To test the hypothesis that the nonlinearity of the Injury Severity Score (ISS) in the log odds of death is responsible for the poor calibration of the TRISS model.
- To evaluate if adding a quadratic ISS term can correct this nonlinearity and improve TRISS model calibration.
Main Methods:
- Examined the linearity of ISS in the log odds of mortality using three trauma datasets: the National Pediatric Trauma Registry (NPTR), University of New Mexico (UNM), and Portland, Oregon (PORT).
- Compared the calibration of TRISS models with and without a quadratic ISS term in the NPTR dataset.
- Assessed model calibration using the Hosmer-Lemeshow statistic and model discrimination.
Main Results:
- ISS demonstrated profound nonlinearity in the log odds of death for both blunt and penetrating trauma in the NPTR dataset (p < 0.001).
- Adding a quadratic ISS term significantly improved TRISS model calibration in the NPTR, evidenced by a 70% reduction in the Hosmer-Lemeshow statistic.
- The addition of the ISS squared term did not impact model discrimination, and similar improvements in log odds of survival modeling were observed in the UNM and PORT datasets.
Conclusions:
- The current TRISS model utilizes ISS in a mathematically inappropriate manner, impairing prediction calibration.
- Incorporating a quadratic ISS term is essential for maintaining TRISS as a reliable prediction standard for trauma.
- Future trauma outcome prediction models require rigorous statistical evaluation, and alternative injury severity scores may necessitate different modeling approaches.