A NEW METHOD FOR PREDICTING SURVIVAL AND ESTIMATING UNCERTAINTY IN TRAUMA PATIENTS

Abstract

The Trauma and Injury Severity Score (TRISS) is the current “gold” standard of screening patient’s condition for purposes of predicting survival probability. More than 40 years of TRISS practice revealed a number of problems, particularly, 1) unexplained fluctuation of predicted values caused by aggregation of screening tests, and 2) low accuracy of uncertainty intervals estimations. We developed a new method made it available for practitioners as a web calculator to reduce negative effect of factors given above. The method involves Bayesian methodology of statistical inference which, being computationally expensive, in theory provides most accurate predictions. We implemented and tested this approach on a data set including 571,148 patients registered in the US National Trauma Data Bank (NTDB) with 1–20 injuries. These patients were distributed over the following categories: (1) 174,647 with 1 injury, (2) 381,137 with 2–10 injuries, and (3) 15,364 with 11–20 injuries. Survival rates in each category were 0.977, 0.953, and 0.831, respectively. The proposed method has improved prediction accuracy by 0.04%, 0.36%, and 3.64% (p-value <0.05) in the categories 1, 2, and 3, respectively. Hosmer-Lemeshow statistics showed a significant improvement of the new model calibration. The uncertainty 2σ intervals were reduced from 0.628 to 0.569 for patients of the second category and from 1.227 to 0.930 for patients of the third category, both with p-value <0.005. The new method shows the statistically significant improvement (p-value <0.05) in accuracy of predicting survival and estimating the uncertainty intervals. The largest improvement has been achieved for patients with 11–20 injuries. The method is available for practitioners as a web calculator http://www.traumacalc.org.