IBM WATSON Trauma Pathway Explorer outperforms the TRISS score to predict early death after polytrauma

Abstract

Abstract Objective In recent years, several big data-based artificial intelligence (AI) systems have found its way in health care, one of which we present here: The IBM WATSON Trauma Pathway Explorer, a visual analytics tool to predict early death in polytrauma patients. The aim of this study was to compare the predictive performance of the Trauma Pathway Explorer for early in-hospital mortality with an established trauma scoring system, the Trauma Revised Injury Severity Score (TRISS). Methods A retrospective comparative accuracy study in a level I trauma center including patients with an Injury Severity Score (ISS) ≥16 and age ≥16 was performed. The compared outcome was early death within 72 hours since admission of the patient. The area under the receiver operating characteristic curve (AUC) was used to measure discrimination. Hosmer-Lemeshow statistics was calculated to analyse calibration of the two predictive models. The Brier score assessed the overall performance of the two models. Results The cohort included 107 polytrauma patients with a death rate of 10.3% at 72 hours since patient admission. The Trauma Pathway Explorer and TRISS showed similar AUCs to predict early death (AUC 0.90, 95% CI 0.79-0.99 vs. AUC 0.88, 95% 0.77-0.97; p = 0.75). The calibration of the Trauma Pathway Explorer was superior to that of TRISS (chi-squared 8.19, Hosmer-Lemeshow p = 0.42 vs. chi-squared 31.93, Hosmer-Lemeshow p < 0.05). The Trauma Pathway Explorer had a lower Brier score than TRISS (0.06 vs. 0.11). Conclusion The IBM WATSON Trauma Pathway Explorer showed equal results in discrimination as TRISS but outperformed in calibration. In addition to being able to provide a prediction of early death, this visual analytics tool for polytrauma patients can also show the quantitative flow of patient sub-cohorts through different events, such as coagulopathy, hemorrhagic shock class, surgical strategy and the above-mentioned outcome. Here, we can present an accurate and valid alternative to TRISS for predicting early death in polytrauma patients.