SSVEP-Based Machine Learning Solution to Classify Mild Traumatic Brain Injury

An objective diagnostic solution for mild traumatic brain injury (mTBI) has been a challenge in sports. The current standard diagnostic tools are considered unreliable due to their reliance on sign and symptom reporting which results in subjective outputs. There is an inherent risk that a mTBI may go unrecognised, which might have a long-term impact on an individual’s cognitive, emotional, and/or motor functions and therefore quality of life. We aimed to develop a solution which provides an objective mTBI assessment. A total of 516 Steady State Visual Evoked Potential (SSVEP) measurements were included in this study, collected from 425 non-mTBI participants and 91 mTBI participants assessed within 72 hours of the injury. Participants were categorised as mTBI following clinical diagnosis by an experienced physician. One minute of SSVEP data was collected from participants using three occipital electrodes using the Nurochek device. Multiple supervised machine learning algorithms were implemented to perform binary classification using various forms of transformed signals, including Fast Fourier Transformation, time-frequency decomposition, and demodulation. The reported results were from stratified k-fold cross-validation, where ‘k’ was the total number of mTBI in the training set. The best cross-validation performance was from Support Vector Machine with 86.19% sensitivity and 60.78% specificity. Using the same classifier on an independent testing set, a sensitivity of 82.00% and specificity of 64.89% were achieved. This research showed the benefits of utilising SSVEP measurements as potential biomarkers for diagnosing acute mTBI. The method and classifier developed could be implemented as a potential diagnostic tool.