Introduction: Motor vehicle collisions (MVC) account for 30-40,000 deaths and 29% of emergency department visits annually in the United States. We recently applied machine-learning tools for the first time to predict clinically meaningful outcomes following MVC.
Methods: We queried two prospectively collected databases maintained at our institution's level one trauma center: i) The American College of Surgeons National Trauma Data Bank (local sample), and ii) the our trauma center's in-house database. De-identified case records were included for all patients who presented following automobile collisions (i.e. excluding motorcycle, bicycle, etc.) and were listed as vehicle occupants (i.e. excluding pedestrian hit by car). Patients were further categorized by mortality and hospital admission. A convolutional neural network (CNN) was trained to predict clinical outcomes and its performance was evaluated.
Results: A total of 17,088 cases were included in our study. Our CNN-derived model predicted mortality at presentation with a 96.8% sensitivity, 92.7% specificity, 92.8% positive predictive value (PPV), 96.7% negative predictive value (NPV), and an overall 98% area under receiver-operator curve (AUROC) from age, Glasgow Coma Scale (GCS), and Injury Severity Score (ISS) alone. (Figure 1) In the sub-group of 16,287 cases who survived and were admitted, our model predicted eventual mortality with 92.1% sensitivity, 90.1% specificity, 90.2% PPV, 92.0% NPV, and an overall 97% AUROC, from age, GCS, ISS, and emergency department length of stay (LOS) alone. (Figure 2) Our model predicted overall hospital LOS in these patients with a mean absolute error of ±4.23 days from the same variables. (Figure 3)
Conclusions: Our CNN-derived models predict clinical outcomes following MVC trauma with high accuracy. This demonstrates the first application of machine-learning to the prediction of MVC clinical outcomes.
Patient Care: By enabling physicians and surgeons treating trauma victims to accurately anticipate clinical outcomes following MVC.
Learning Objectives: At the conclusion of this session, participants should be able to:
1) Describe key predictors of mortality after MVC
2) Use our model to predict morbidity and hospital LOS after MVC
3) Discuss the proper application of machine learning models to clinical decision-making
References: 1. Florence C, Haegerich T, Simon T, Zhou C, Luo F. Estimated Lifetime Medical and Work-Loss Costs of Emergency Department-Treated Nonfatal Injuries--United States, 2013. MMWR. Morbidity and mortality weekly report. 2015;64(38):1078-1082.
2. Web-Based Injury Statistics Query and Reporting System (WISQARS). Centers for Disease Control and Prevention, National Center for Injury Prevention and Control;2016.
3. Blincoe LJ MT, Zaloshnja E, Lawrence BA. The economic and societal impact of motor vehicle crashes, 2010 (revised). National Highway Traffic Safety Administration;2015. DOT HS 812 013.
4. Rating the severity of tissue damage. I. The abbreviated scale. Jama. 1971;215(2):277-280.
5. Baker SP, O'Neill B, Haddon W, Jr., Long WB. The injury severity score: a method for describing patients with multiple injuries and evaluating emergency care. The Journal of trauma. 1974;14(3):187-196.