D C Becalick1, T J Coats. 1. Academic Unit of Accident and Emergency, St. Bartholomew's and the Royal London School of Medicine, Queen Mary and Westfield College, London, England.
Abstract
BACKGROUND: The development of TRISS was principally a search for variables that correlated with outcome. It is not known, however, if linear statistical models provide optimal results. Artificial intelligence techniques can answer this question and also determine the most important predictor variables. METHODS: An artificial neural network, using 16 anatomic and physiologic predictor variables, was compared with the latest United Kingdom version of TRISS model. RESULTS: Both methods were 89.6% correct, but TRISS was significantly better by the area under the receiver operating characteristic curve (0.941 vs. 0.921, p < 0.001). The artificial neural network, however, was better calibrated to the test data (Hosmer-Lemeshow statistic, 58.3 vs. 105.4). Head injury, age, and chest injury were the most important predictors by linear or nonlinear methods, whereas respiration rate, heart rate, and systolic blood pressure were underused. CONCLUSION: Prediction using linear statistics is adequate but not optimal. Only half the predictors have important predictive value, fewer still when using linear classification. The strongest predictors swamp any nonlinearity observed in other variables.
BACKGROUND: The development of TRISS was principally a search for variables that correlated with outcome. It is not known, however, if linear statistical models provide optimal results. Artificial intelligence techniques can answer this question and also determine the most important predictor variables. METHODS: An artificial neural network, using 16 anatomic and physiologic predictor variables, was compared with the latest United Kingdom version of TRISS model. RESULTS: Both methods were 89.6% correct, but TRISS was significantly better by the area under the receiver operating characteristic curve (0.941 vs. 0.921, p < 0.001). The artificial neural network, however, was better calibrated to the test data (Hosmer-Lemeshow statistic, 58.3 vs. 105.4). Head injury, age, and chest injury were the most important predictors by linear or nonlinear methods, whereas respiration rate, heart rate, and systolic blood pressure were underused. CONCLUSION: Prediction using linear statistics is adequate but not optimal. Only half the predictors have important predictive value, fewer still when using linear classification. The strongest predictors swamp any nonlinearity observed in other variables.
Authors: Ashish Nimgaonkar; Dilip R Karnad; S Sudarshan; Lucila Ohno-Machado; Isaac Kohane Journal: Intensive Care Med Date: 2004-01-15 Impact factor: 17.440
Authors: Hideo Tohira; Ian Jacobs; David Mountain; Nick Gibson; Allen Yeo Journal: Scand J Trauma Resusc Emerg Med Date: 2012-09-10 Impact factor: 2.953