Interpreting EEG Signals with Concept-Based Explainable AI

The global shortage of neurologists, particularly in under-resourced regions, hinders the early diagnosis of neurological disorders. Electroencephalography (EEG), a non-invasive method for recording brain activity, is widely used for diagnosis but its interpretation is a timeconsuming process that could delay patient care. Deep learning models have been used for EEG classification, but their black-box nature limits their clinical adoption. Explainable AI (XAI) techniques attempt to address this, yet most rely on feature-attribution approaches that provide explanations on low-level features rather than aligning with abstract neurological concepts. This work aims to explain the EEG classification model by verifying that the model learns and uses relevant concepts as used by neurologists. We employ Testing with Concept Activation Vectors (TCAV) for global explanations, quantifying the influence of pathological EEG events such as spike and slow waves, generalized periodic epileptiform discharges, periodic lateralized epileptiform discharges and low-frequency waveforms. For local explanations, conceptual sensitivity analysis measures the influence of these concepts on individual EEGs. Experiments on publicly available Temple University Hospital (TUH) EEG dataset confirm that abnormal class prediction is sensitive to pathological events and low-frequency bands, while non-pathological events such as eye movement, artifact and background activity are neutral concepts. The alignment between global TCAV scores and local conceptual sensitivity improves the model interpretability and will largely facilitate its integration into clinical workflows. This work is the first to extend TCAV-based conceptual sensitivity for generating local, per-sample concept-based explanations in EEG classification.