Transformed wavelets for motor imagery EEG classification using hybrid CNN-modified vision transformer: an exploratory study of MI EEG

Wavelets capture signal characteristics across time and frequency, but traditional wavelets suffer from high time–bandwidth products (TBP), limiting feature discrimination in EEG classification. We propose transformed wavelets with improved TBP and frequency bandwidth, outperforming Morlet by 0.04 and 0.20, respectively. Using datasets BCI Competition IV 2a, 2b, and CLA, we evaluated both fundamental and transformed wavelets with a modified vision transformer (MViT). Enhanced scalograms generated through local mean and principal component analysis (PCA) consistently outperformed raw scalograms. A hybrid convolutional neural network (CNN)–MViT achieved 82.35% inter-subject and 89.02% intra-subject accuracy, with 3–4% average gains in motor imagery EEG decoding.