Explainable AI to Identify Feature Importance in EEG Data for Mild Traumatic Brain Injury

Mild Traumatic Brain Injury (mTBI) is a prevalent yet often underdiagnosed condition, with many patients failing to receive structured follow-up care. While CT scans remain the standard for detecting severe brain injuries, they are often ineffective in mild cases exposing patients to unnecessary radiation. Electroencephalography (EEG) is a promising prediagnostic tool for mTBI detection by capturing brain wave abnormalities. In this study, we utilize an LSTM-based deep learning based model to classify patients as CT positive or CT negative based on EEG data. However, deep learning models, despite their high accuracy, often function as “black boxes,” making it difficult to interpret which features contribute most to their predictions. To enhance interpretability, we leverage Explainable AI (XAI) techniques using Captum’s FeaturePermutation to identify the most influential electrodes and clinical features. Our findings reveal that central and midline electrodes (Pz, Fz, Cz), along with key clinical factors such as the Glasgow Coma Scale (GCS) and time difference between time of injury to time of recording, significantly contribute to classification performance. The model achieved an impressive accuracy of 99.96%, highlighting the effectiveness of EEG-based assessment. By ranking features based on importance, we demonstrate the potential for reducing feature dimensionality while maintaining accuracy. These insights not only improve model transparency but also pave the way for future research on EEG-based biomarkers for brain injury detection and personalized clinical interventions.