BACKGROUND: The Trauma and Injury Severity Score (TRISS) has been the approach to trauma outcome prediction during the past 20 years and has been adopted by many commercial registries. Unfortunately, its survival predictions are based upon coefficients that were derived from a data set collected in the 1980s and updated only once using a data set collected in the early 1990s. We hypothesized that the improvements in trauma care during the past 20 years would lead to improved survival in a large database, thus making the TRISS biased. METHODS: The TRISSs from the Pennsylvania statewide trauma registry (Collector, Digital Innovations) for the years 1990 to 2010. Observed-to-expected mortality ratios for each year of the study were calculated by taking the ratio of actual deaths (observed deaths, O) to the summation of the probability of mortality predicted by the TRISS taken over all patients (expected deaths, E). For reference, O/E ratio should approach 1 if the TRISS is well calibrated (i.e., has predictive accuracy). RESULTS: There were 408,489 patients with complete data sufficient to calculate the TRISSs. There was a significant trend toward improved outcome (i.e., decreasing O/E ratio; nonparametric test of trend, p < 0.001) over time in both the total population and the blunt trauma subpopulation. In the penetrating trauma population, there was a trend toward improved outcome (decreasing O/E ratio), but it did not quite reach significance (nonparametric test of trend p = 0.073). CONCLUSION: There is a steady trend toward improved O/E survival in the Pennsylvania database with each passing year, suggesting that the TRISS is drifting out of calibration. It is likely that improvements in care account for these changes. For the TRISS to remain an accurate outcome prediction model, new coefficients would need to be calculated periodically to keep up with trends in trauma care. This requirement for occasional updating is likely to be a requirement of any trauma prediction model, but because many other deficiencies in the TRISS have been reported, we think that rather than updating the TRISS, it would be more productive to replace the TRISS with a modern statistical model.
BACKGROUND: The Trauma and Injury Severity Score (TRISS) has been the approach to trauma outcome prediction during the past 20 years and has been adopted by many commercial registries. Unfortunately, its survival predictions are based upon coefficients that were derived from a data set collected in the 1980s and updated only once using a data set collected in the early 1990s. We hypothesized that the improvements in trauma care during the past 20 years would lead to improved survival in a large database, thus making the TRISS biased. METHODS: The TRISSs from the Pennsylvania statewide trauma registry (Collector, Digital Innovations) for the years 1990 to 2010. Observed-to-expected mortality ratios for each year of the study were calculated by taking the ratio of actual deaths (observed deaths, O) to the summation of the probability of mortality predicted by the TRISS taken over all patients (expected deaths, E). For reference, O/E ratio should approach 1 if the TRISS is well calibrated (i.e., has predictive accuracy). RESULTS: There were 408,489 patients with complete data sufficient to calculate the TRISSs. There was a significant trend toward improved outcome (i.e., decreasing O/E ratio; nonparametric test of trend, p < 0.001) over time in both the total population and the blunt trauma subpopulation. In the penetrating trauma population, there was a trend toward improved outcome (decreasing O/E ratio), but it did not quite reach significance (nonparametric test of trend p = 0.073). CONCLUSION: There is a steady trend toward improved O/E survival in the Pennsylvania database with each passing year, suggesting that the TRISS is drifting out of calibration. It is likely that improvements in care account for these changes. For the TRISS to remain an accurate outcome prediction model, new coefficients would need to be calculated periodically to keep up with trends in trauma care. This requirement for occasional updating is likely to be a requirement of any trauma prediction model, but because many other deficiencies in the TRISS have been reported, we think that rather than updating the TRISS, it would be more productive to replace the TRISS with a modern statistical model.
Authors: Rachael A Callcut; Glenn Wakam; Amanda S Conroy; Lucy Z Kornblith; Benjamin M Howard; Eric M Campion; Mary F Nelson; Matthew W Mell; Mitchell J Cohen Journal: J Trauma Acute Care Surg Date: 2016-02 Impact factor: 3.313
Authors: Kei Ching Kevin Hung; Chun Yu Lai; Janice Hiu Hung Yeung; Marc Maegele; Po Shan Lily Chan; Ming Leung; Hay Tai Wong; John Kit Shing Wong; Ling Yan Leung; Marc Chong; Chi Hung Cheng; Nai Kwong Cheung; Colin Alexander Graham Journal: Eur J Trauma Emerg Surg Date: 2021-04-26 Impact factor: 3.693
Authors: Amy C Gunning; Koen W W Lansink; Karlijn J P van Wessem; Zsolt J Balogh; Frederick P Rivara; Ronald V Maier; Luke P H Leenen Journal: World J Surg Date: 2015-11 Impact factor: 3.352