Development and Validation of Two Prediction Models for 72-Hour Mortality in High-Risk Trauma Patients Using a Benchmark Dataset: A Comparative Study of Logistic Regression and Neural Networks Models

Background Many studies have been conducted to develop scoring systems for trauma patients, with the majority using logistic regression (LR) models. Neural networks (NN), which is a machine learning algorithm, has a potential to increase the performance of these models. Objectives The aim of this study was to develop and validate two separate prediction models for 72-hour mortality of high-risk trauma patients using LR and NN and to compare the performances of these models in detail. We also aimed to share the SPSS calculators for our models. Materials and methods This is a retrospective, single-center study conducted using a benchmark dataset where the patients were retrospectively gathered from a level 1 trauma center. Patients older than 18 years of age, who had multiple injuries, and were treated at the University Hospital Zurich between January 1, 1996, and January 1, 2013, were included. Patients with a condition that may have an impact on the musculoskeletal system, with Injury Severity Score<16, and with missing outcome data were excluded. Results A total of 3,075 patients were included in the analysis. The area under the curve values of the LR and NN models for predicting 72-hour mortality in patients with high-risk trauma in the hold-out cohort were 0.859 (95% CI=0.836 to 0.883) and 0.856 (95% CI=0.831 to 0.880), respectively. There was no statistically significant difference in the performance of the models (p = 0.554, DeLong's test). Conclusion Both of the models showed good discrimination. Our study suggests that the NN and LR models we developed hold promise as screening tools for predicting 72-hour mortality in high-risk trauma patients. These models were made available to clinicians as clinical prediction tools via SPSS calculators. However, further external validation studies in diverse populations are necessary to substantiate their clinical utility. Moreover, in subsequent studies, it would be beneficial to derive NN models with substantial events per predictor variable to attain more robust and greater predictive accuracy. If the dataset is relatively limited, using LR seems to be a viable alternative.