Accuracy of algorithms to predict injury severity in older adults for trauma triage

• Published in: Traffic injury prevention 2019-11, Vol.20 (sup2), p.S81-S87
• Main Authors: Hartka, Thomas, Gancayco, Christina, McMurry, Timothy, Robson, Marina, Weaver, Ashley
• Format: Article
• Language: English
• Subjects: Accuracy; Adults; Age; Algorithms; Crashworthiness; Fragility; Health risks; Impact strength; Injuries; Injury analysis; Injury outcome; injury prediction; Model accuracy; Motor vehicles; NASS-CDS; Occupant injuries; older occupant; Older people; Prediction models; Regression analysis; Rollover; Traffic accidents & safety; Trauma; trauma triage
• ISSN: 1538-9588; 1538-957X
• DOI: 10.1080/15389588.2019.1688795

Objective: Older adults make up a rapidly increasing proportion of motor vehicle occupants and previous studies have demonstrated that this population is more susceptible to traumatic injuries. The CDC recommends that patients anticipated to have severe injuries (Injury Severity Score [ISS] ≥ 16) be transported to a trauma center. The recommended target rate for undertriage is ≤ 5% and for overtriage is ≤ 50%. Several regression-based algorithms for injury prediction have been developed in order to predict severe injury in occupants involved in a motor vehicle collision (MVC). The objective of this study to was to determine if the accuracy of regression-based injury severity prediction algorithms decreases for older adults. Methods: Data were obtained from the National Automotive Sampling System - Crashworthiness Data System (NASS-CDS) from the years 2000-2015. Adult occupants involved in non-rollover MVCs were included. Regression-based injury risk models to predict severe injury (ISS ≥ 16) were developed using random split-samples with the following variables: age, delta-V, direction of impact, belt status, and number of impacts. Separate models were trained using data from the following age groups: (1) all adults, (2) 15-54 years, (3) ≥45 years, (4) ≥55 years, and (5) ≥65 years. The models were compared using the mean receiver operating characteristic area under curve (ROC-AUC) after 1,000 iterations of training and testing. The predicted rates of overtriage were then determined for each group in order to achieve an undertriage rate of 5%. Results: There were 24,577 occupants (6,863,306 weighted) included in this analysis. The injury prediction model trained using data from all adults did not perform as well when tested on older adults (ROC-AUC: 15-54 years: 0.874 [95% CI: [0.851-0.895]; 45+ years: 0.837 [95% CI: 0.802-869]; 55+ years: 0.821 [95% CI: 0.775-0.864]; and 65+ years: 0.813 [95% CI: 0.754-0.866]). The accuracy of this model decreased in each decade of life. The performance did not change significantly when age-specific data were used to train the prediction models (ROC-AUC: 18-54 years: 0.874 [95% CI: 0.851-0.896]; 45+ years: 0.836 [95% CI: 0.798-0.871]; 55+ years: 0.822 [95% CI: 0.779-0.864]; and 65+ years: 0.808 [95% CI: 0.748-0.868]). In order to achieve an undertriage rate of 5%, the predicted overtriage rate by these models were 50% for occupants 15-54 years, 61% for occupants ≥ 55 years, 70% for occupants ≥ 55 years, and 71% for occupan