Linking a Deep Learning Model for Concussion Classification with Reorganization of Large-Scale Brain Networks in Female Youth.

Abstract

Concussion, or mild traumatic brain injury, is a significant public health challenge, with females experiencing high rates and prolonged symptoms. Reliable and objective tools for early diagnosis are critically needed, particularly in pediatric populations, where subjective symptom reporting can be inconsistent and neurodevelopmental factors may influence presentation. Five minutes of resting-state (RS) EEG data were collected from non-concussed and concussed females between 15 and 24 years of age. We first applied a deep learning approach to classify concussion directly from raw, RS electroencephalography (EEG) data. A long short-term memory (LSTM) recurrent neural network trained on the raw data achieved 84.2% accuracy and an ensemble median area under the receiver operating characteristic curve (AUC) of 0.904. To complement these results, we examined causal connectivity at the source level using information flow rate to explore potential network-level changes associated with concussion. Effective connectivity in the non-concussed cohort was characterized by a symmetric pattern along the central-parietal midline; in contrast, the concussed group showed a more posterior and left-lateralized pattern. These spatial distribution changes were accompanied by significantly higher connection magnitudes in the concussed group (p < 0.001). While these connectivity changes may not directly drive classification, they provide evidence of large-scale brain reorganization following concussion. Together, our results suggest that deep learning models can detect concussion with high accuracy, while connectivity analyses may offer complementary mechanistic insights. Future work with larger datasets is necessary to refine the model specificity, explore subgroup differences related to hormone cycle changes and symptoms, and incorporate data across different sports.