EEG-CLT: A CNN-sLSTM-Transformer sequential architecture for classification of motor imagery from electroencephalogram signals

Motor imagery (MI)-based brain–computer interfaces (BCIs) using electroencephalogram (EEG) signals have received particular attention due to their potential to assist in motor function rehabilitation and communication support, thereby improving the quality of life for individuals with physical impairments. MI-BCI classification methods based on deep learning architectures have gained significant popularity in recent years. To further enhance classification performance, we propose EEG-CLT, a sequential architecture that combines three deep learning modules: a Convolutional Neural Network (CNN) to extract local spatiotemporal features, scalar Long Short-Term Memory (sLSTM) networks to capture temporal dependencies, and a Transformer for integrating global context through the self-attention mechanism. The proposed model achieved robust classification performance in within-subject and cross-subject evaluations. In the within-subject evaluation, the accuracies were 75.8% and 87.1% on the BCI Competition IV 2a and 2b datasets, respectively. In the cross-subject evaluation, the accuracies were 61.7%, 74.4%, and 63.1% on the BCI Competition IV 2a, 2b, and PhysioNet EEG datasets, respectively. These accuracies were higher than or comparable to those of the baseline deep learning models. These results suggest that EEG-CLT, which incorporates sLSTM into a CNN–Transformer-based sequential pipeline, provides a useful and competitive design choice for MI-EEG classification across multiple public datasets.