A machine learning based approach for identifying traumatic brain injury patients for whom a head CT scan can be avoided.

Head CT scan is more often used to evaluate patients with suspected traumatic brain injury (TBI). However, the use of head CT scans in evaluating TBI is costly with low value endeavor. In this paper, we propose a new algorithm and a set of features to help clinicians determine which patients evaluated for TBI need a head CT scan using cost sensitive random forest (CSRF) classifier. We show that random forest (RF) and CSRF are useful methods for identifying patients likely to have a positive head CT scan. The proposed algorithm has superior diagnostic accuracy in comparison to the Canadian head CT algorithm, which is currently the most accurate and widely used algorithm for determining which TBI patients need a head CT scan. In the highest sensitivity (i.e. 100%), our method outperforms the Canadian rule in terms of specificity, accuracy and area under ROC curve using cost sensitive classifier. Clinical implementation of this algorithm can help decrease financial costs associated with Emergency Department evaluations for traumatic brain injury, while decreasing patient exposure to avoidable ionizing radiation.