Advancing Motor Imagery EEG Classification through Wavelet Scattering Transforms and 1D Transformers

Motor imagery (MI) electroencephalography (EEG) classification is critical to brain-computer interfaces (BCI). Existing techniques failed mainly because of the noise and non-linearity features of EEG signals. This research work develops a new approach that exploits wavelet scattering transform (WST) for feature extraction along with the 1D transformer models for classification. WST captures time-frequency information and preserves locality of the signal; the 1D transformer uses its self-attention mechanism to take advantage of sequential data and capture long-range dependencies, strengthening the model’s capacity to focus only on relevant features in the signal. The proposed method achieves an accuracy of 98.10% on a well-known EEG dataset, surpassing the performance of several existing methods. The model was evaluated on other metrics like precision, recall, and F1 score, giving an average value of 98%. This work further advances the MI EEG classification, and more robust and accurate BCI systems can be laid by making use of the best of both WST and transformer architectures.