EEG Channel Selection Based on Time–Frequency Hellinger Distance

This paper presents a novel method for EEG channel selection based on the Hellinger Distance (HD) computed over time–frequency representations (TFRs). Here, we first convert raw EEG into the time–frequency domain using Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT). Leveraging Hellinger Distance, we then identify and remove outlier epochs via a z-score threshold, and rank and select the most discriminative channels by measuring how well each channel’s TFR distributions separate different classes. Our empirical evaluation uses the BCI Competition IVa dataset to compare the proposed HD-based approach (HD-CSP) against multiple variants of the Common Spatial Pattern algorithm, including standard CSP, L1 Norm CSP, SCSP, FBCSP, and E-CSP. Results indicate that HD-CSP consistently outperforms competing methods in all tested configurations, achieving notably high classification accuracy even when the number of channels is severely restricted. In particular, HD-CSP reaches around 70\% accuracy with only three channels, while other approaches suffer significant performance drops. As the number of channels increases, HD-CSP maintains its superior accuracy, exceeding 80% in some configurations. Overall, the proposed method is superior on performance gains and ability to adapt to diverse channel configurations suggest broad applicability, especially in resource-constrained EEG settings where efficiency and accuracy are both priorities.