Insights Into Pre-Stimulus Activity for EEG-Based Machine Learning Sensory Classification

Understanding how the brain perceives the world is a central aim in neuroscience, often investigated through Electroencephalography (EEG). Traditional neuroscience methodologies have enriched our understanding of sensory processing, but their integration with Machine Learning (ML) remains limited. This study combines ML with EEG techniques to classify auditory and visual Event-related potentials (ERPs). Our investigation focuses on evaluating the impact of both pre-stimulus and post-stimulus ERPs components on model performance. We trained eXtreme Gradient Boosting (XGBoost) to classify auditory and visual ERPs using either a short window with only post-stimulus components or a long window with both pre-stimulus and post-stimulus components, then compared the model performance to a Support Vector Machine (SVM) as a baseline model. We employed an interpretable AI approach to examine the global feature importance by using feature importance of XGBoost, and SHapley Additive exPlanations (SHAP) for the SVM. XGBoost achieved an accuracy of 86% for both pre-post-stimulus and post-stimulus components, outperforming the SVM by 7 and 2 percentage points, respectively. The analysis of global feature importance revealed XGBoost's effectiveness in focusing on post-stimulus peaks and pre-stimulus components, aspects that the baseline model did not sufficiently capture. These findings give insights on how to use ERPs data when approaching sensory classification problems.