Prognosticating outcome using magnetic resonance imaging in patients with moderate to severe traumatic brain injury: a machine learning approach

Introduction Over the last decade advancements in computer processing have enabled the application of machine learning (ML) to complex medical problems. Convolutional neural networks (CNN), a type of ML, have been used to interrogate medical images for variety of purposes. In this study, we aimed to investigate the potential application of CNN in prognosticating patients with traumatic brain injury (TBI). Methods Patients with moderate to severe TBI and evidence of diffuse axonal injury (DAI) were selected retrospectively. A CNN model was developed using a training subgroup and a holdout subgroup was used as a testing dataset. We reported the model characteristics including area under the receiver operating characteristic curve (AUC). Results We included a total of 38 patient, of which we generated 725 MRI sections. We developed a CNN model based on a modified AlexNet architecture that interpreted the brain stem injury to generate outcome predictions. The model was able to predict GOS outcomes with a specificity of 0.43 and a sensitivity of 0.997. It showed an AUC of 0.917. Conclusion The utilization of machine learning MRI analysis for prognosticating patients with TBI is a valued method that require further investigation. This will require multicentre collaboration to generate large datasets.

Automated summary

This study aimed to investigate the potential application of machine learning and convolutional neural networks (CNN) in prognosticating patients with traumatic brain injury (TBI). The results showed that the developed CNN model was able to predict the prognosis of patients by interpreting brain stem injuries.