Investigating the Effect of Pre-processing Methods on Model Decision-Making in EEG-Based Person Identification

Electroencephalography (EEG) data has emerged as a promising modality for biometric applications, offering unique and secure personal identification and authentication methods. This research comprehensively compared EEG data pre-processing techniques, focusing on biometric applications. In tandem with this, the study illuminates the pivotal role of Explainable Artificial Intelligence (XAI) in enhancing the transparency and interpretability of machine learning models. Notably, integrating XAI methodologies contributes significantly to the evolution of more precise, reliable, and ethically sound machine learning systems. An outstanding test accuracy exceeding 99% was observed within the biometric system, corroborating the Graph Neural Network (GNN) model's ability to distinguish between individuals. However: high accuracy does not unequivocally signify that models have extracted meaningful features from the EEG data. Despite impressive test accuracy, a fundamental need remains for an in-depth comprehension of the models. Attributions proffer initial insights into the decision-making process. Still, they did not allow us to determine why specific channels are more contributory than others and whether the models have discerned genuine cognitive processing discrepancies. Nevertheless, deploying explainability techniques has amplified system-wide interpretability and revealed that models learned to identify noise patterns to distinguish between individuals. Applying XAI techniques and fostering interdisciplinary partnerships that blend the domain expertise from neuroscience and machine learning is necessary to interpret attributions further and illuminate the models' decision-making processes.