Enhancing EEG based cognitive state classification using graph Fourier transform

Electroencephalography (EEG) based cognitive state classification has been widely explored with deep-learning models demonstrating remarkable performance. Deep-learning models need high computational resources and large datasets, creating a need for an alternative methodology. Graph signal processing (GSP) techniques provide an effective lightweight alternative by capturing spatial dependencies across the channels. This study investigates the effectiveness of GSP-based graph Fourier transform (GFT) features on a publicly available EEG mental arithmetic task (EEGMAT) dataset. A strict 5-fold cross-validation has been used as an evaluation technique. GFT is applied to extract spatial-spectral features by modeling EEG-channels as graph nodes. The extracted features have been evaluated using multiple classifiers including Random Forest (RF), Extreme Gradient Boosting (XGB), Decision Tree (DT), and Logistic Regression (LR). Statistical analysis confirms the significance of GFT features compared to raw signals. RF gave the highest accuracy of approximately 99 percent. Model interpretability based on Shapley additive explanations (SHAP) revealed that the frontal and central regions contributed to the classification aligning with the findings of cognitive neuroscientists.