Student Presenter(s): Scott Kivitz, Dana Schulz, Stephanie De Mel, Sonia Amanat, Taner Celebi
Faculty Mentor: Stephen DiRusso
Department: Clinical Specialties
School/College: College of Osteopathic Medicine, Long Island

Introduction: Current practice is to use externally validated survival prediction models to assess mortality. In this study, we use a locally generated prediction model that is internally validated to assess that Trauma Service's performance over time.

Methods: Retrospective study design used an urban Level-II ACS Trauma Center trauma registry (2016 to 2019). Mortality prediction was modeled using Multilayer Perception Artificial Neural Network (ANN). Independent Variable Importance was calculated. Discrimination (Area under the Receiver Operator Curve (AuROC)) and calibration (Hosmer-Lemeshow C-statistic (HL-C)) measured predictive capability.

Results: Model included 3,468 patients, 2,581 (74.4%) in 2016–2018 (training set), and 887 (25.6%) in 2019 (test set), with AuROC of 0.947 and an HL-C of 12.75 (p>0.05). Predicted and observed mortality (95% CI) from 2016–2018 was 3.56% (3.11, 4.01) and 4.0% (3.21, 4.78), respectively. With 2016–2018 as a reference, 2019 predicted mortality (3.65%, 95% CI: 2.91, 4.41) did not differ from observed mortality (3.0%, 95% CI: 1.82, 4.18). Although not statistically significant, the trauma service improved its mortality rate by 25% (2016–2018 compared to 2019).

Conclusions: We generated and internally validated an ANN model with excellent prediction of survival and used this to assess Trauma Service performance for the subsequent year. We will use this model on a continual basis to benchmark Trauma Service performance.