Mild traumatic brain injuries (mTBIs) are the most common type of brain injury, accounting for the vast majority of head injuries. However, diagnosing mTBI remains a challenge, as clear indicators may not be present in symptom questionnaires or structural imaging modalities, which are not feasible for deployment at the site of the injury. Therefore, this study explores the development of a diagnostic algorithm that incorporates novel sets of biomarkers derived from noninvasive sensors, specifically electroencephalography (EEG) and electrocardiogram (ECG). Using a publicly available dataset, we explore the relationship between sensor data, symptom questionnaires, and clinical diagnoses. We extract temporal, spectral, and spatial domain features from the sensor data, leveraging techniques from network science and multi- variate time-series analysis. We evaluate an ensemble of linear and nonlinear machine- learning classifiers on unseen holdout data. Integrating sensor data with symptom questionnaires yields over 0.90 area under the receiver operating characteristic curve (AUROC), demonstrating utility as a diagnostic aid. Furthermore, the EEG features alone achieve above 0.70 AUROC, outperforming baseline EEG-only approaches on this dataset. As a result, this work advances mTBI diagnostic knowledge by estab- lishing a benchmark classification performance and highlighting potential biomarkers of cognitive or neurological dysfunction.

Figures: 

1. Proposed framework for the multimodal detection of mild traumatic brain injury (Top)
2. Summary of the data processing and machine learning steps. Panel a.: Overview of the semi-automated pre-processing procedure with an illustrative example of data containing heartbeat artifacts. Data is re-referenced to the Surface Laplacian montage, highpass filtered above 0.3Hz, and line-noise and heartbeat artifacts are removed. Data is segmented into eyes open and eyes closed segments free of artifacts. Panel b. Examples of the types of features extracted from the EEG data. Panel c. Model development pipeline, including feature selection, machine learning model training, nested cross-validation for hyperparameter tuning, and final metalearning model selection and model stacking. (Bottom)