Injury Risk Prediction for Body Regions after Motor Vehicle Collisions to Guide CT Scanning Decisions

Abstract

Full body computed tomography (CT) is a widely used clinical evaluation method to detect hidden injury for victims of motor vehicle collisions (MVCs). However, full body CT scans are time consuming and expensive for both healthcare service providers and MVC victims. Injury risk prediction models that support CT scanning decisions are therefore highly desired. Existing studies have implemented logistic regression models to predict injury risk for victims' major body regions, including head, neck, chest, abdomen/pelvis, cervical spine, thoracic spine and lumbar spine. The work presented here involved the application of novel approaches to improve the prediction results. This study focused on examining patient information and crash data for front seat adult passengers using data from the National Automotive Sampling System - Crashworthiness Data System from 2000 to 2015. This dataset is imbued with a large amount of missingness and is highly imbalanced. Various imputation methods were employed in order to preserve the greatest amount of relevant historical data possible. The high imbalance in the data was resolved by the implementation of downsampling and synthetic minority over-sampling technique. Models that were applied in this study include logistic regression, random forests, support vector machines and gradient boosting. Autoencoders were also deployed to generate features of high importance to improve prediction results. The resulting models for all seven regions yielded sensitivities and specificities of at least 96% and 30%, respectively. Overall, these models were developed not to replace physicians' decisions, but to guide their CT scanning decisions.