Background
Traumatic Brain Injury (TBI) is a prevalent condition in the United States (U.S.), affecting approximately 1.7 million individuals annually. Initially thought to cause minor harm, mild TBI (mTBI), or concussion, has gained considerable attention owing to its serious neuropsychological consequences, notably in athletes participating in contact sports and military personnel. The Glasgow Coma Scale, the gold standard for assessing head injury, is a subjective score made by a physician on the patient’s gross motor and verbal responses and is not sensitive to mTBI severity. Electroencephalography (EEG) has shown promise as an objective biomarker for mTBI; however, evaluating raw EEG is time consuming and difficult. Automated approaches implementing frequency domain analyses of EEG have shown modest classification of mTBI, but they may not fully capture TBI pathology. Spiking neural networks (SNNs) are a form of neuromorphic computing, or bio-inspired machine learning, that model time series data with high accuracy and operate at low power. SNNs have the potential to uncover aspects of brain dynamics hidden from traditional frequency domain approaches, and we are testing whether these features are more sensitive to TBI diagnosis and predicting recovery.
Approach
Our research will design and train SNN architectures that model TBI brain dynamics in a public, longitudinal study consisting of mTBI patients and healthy controls. All study participants were assessed at 2 weeks, 2 months, and 4 months post-injury, with EEG and post-concussive symptoms collected at each session. The collected EEG data includes tasks that engage working memory and auditory processing, as well as at rest. After removing mechanical and physiological, non-brain artifacts, SNN models will classify mTBI from healthy EEG at baseline. The features from this model will then be used to assess symptom severity at baseline and recovery months later.
Accomplishments
This project was recently initiated in July 2024. Since then, we have developed a pre-processing pipeline to remove non-brain artifacts, such as eye blinks and muscle responses, from EEG data to ensure downstream machine learning performance is due to brain dynamics and not due to peripheral signals.