Development and Validation of a Computed Tomography Based Artificial Intelligence Screening Tool for Acute Cervical Spine Trauma

Introduction: In the setting of cervical spinal trauma (CST), there is high quality evidence demonstrating earlier intervention leads to improved outcomes. This poses a significant triage challenge to regional hospitals reliant on patient transfer for surgical treatment. In this study, we develop and test an artificial intelligence (AI) based screening tool to predict likelihood of surgical intervention for patients with a traumatic injury to the cervical spinal column and/or spinal cord using computed tomography (CT) imaging obtained in the emergency department.

Methods: Patients with CST treated at a level 1 adult trauma center in Ontario, Canada from 2005 to 2023 were retrospectively included using a local trauma registry. Two channel separated convolutional networks (CSNs), three two-dimensional combination convolutional and recurrent neural networks (2D CNN-RNN), and two vision transformer (ViT) models were trained, internally validated, and tested using cervical spine CT scans. Binary patient-level labels corresponding to whether the patient received surgical intervention to the spine served as the reference standard.

Results: There were 3,068 trauma patients with spine CT scans included in the study. There were 383 patients (12.5%) that underwent surgery for CST. There were 2,254 patients in the training and validation cohort (N=286 underwent surgery, 12.7%), 398 patients in the internal test cohort (N=50 underwent surgery, 12.6%), and 416 patients in the hold-out test cohort (47 underwent surgery, 11.3%). The CSN models were found to have the greatest mean sensitivity (91.5%, 95% CI: 80.1 - 96.6%), specificity (94.0%, 95% CI 91.1 – 96.0%), area under the receiver operating characteristic curve (0.93, 95% CI: 0.89 - 0.97), negative predictive value (98.7%; 95% CI 96.7 – 99.5%), and positive predictive value (61.4%; 95% CI 48.4 – 72.9%).

Conclusion : This study demonstrates, for the first time, that AI-based prediction models can be used to identify patients with CST who are likely to require surgical intervention.