The high incidence of concussions/mild traumatic brain injuries (mTBI) in the United States constitutes a major public health crisis. mTBIs, especially repetitive mTBIs, lead to a wide array of acute and chronic symptoms and health concerns, with annual health care costs in the billions of dollars. Current clinical tools to detect acute and early stage mTBI, tools which mostly rely on imaging or behavioral and symptom-based assessment, have poor to moderate performance. There is thus a need for a fast, accurate, and quantitative method to aid in the identification and management of mTBI. A method combining bioelectrical signal data from low power hardware with recent advances in machine learning and graphical signal processing could augment current clinical approaches. However, despite growing research efforts surrounding quantitative concussion diagnosis, there is still limited agreement on the most useful algorithms and types of bioelectrical features for the prediction of mTBI, and concerns over model interpretability and feature validity remain. Many recently proposed machine learning-based diagnostic methods for mTBI rely heavily on electroencephalography (EEG) features and have seen limited and inconsistent success. On the other hand, features related to autonomic nervous system (ANS) like heart rate variability (HRV), as well as features related to galvanic skin response (GSR), have been understudied for mTBI prediction despite mounting clinical evidence of links between mTBI and changes in HRV and GSR. We use open-source data and clinical data from collaborators at Houston Methodist to identify important features and interpretable machine learning solutions to detect mTBI from short term effects. Exploring a broad range of biological signals and indicators, we develop novel graphical signal processing-based approaches to smartly combine multimodal data and uncover additional, underexplored parameters based on graph properties and structure. If successful, our data driven framework will result in high classification accuracy of mTBI based on short-term bioelectrical signal recordings from portable hardware. Our data processing routines could have use in future smart helmet technologies, and the development of our framework may serve as a first step toward uncovering techniques for signal modulation-based treatment of concussion as well.
Collaborators: Gavin Britz M.D., Taiyun Chi, Eugene Golanov M.D., and Kenneth Podell
Current Researchers: Anton Banta, Roy Phillips