Classifying imagined movements using dynamic multi-scale convolution and multi-head time-focused attention

The proposed model achieves average accuracies of 80.32% and 90.81% on BCI Competition IV datasets 2a and 2b, respectively.

• Convolutional neural networks are effective for classifying motor imagery electroencephalogram signals, but performance varies significantly between individuals.
• Dynamic multi-scale CNN and multi-head temporal attention are combined to enhance feature extraction from MI-EEG signals.
• The model utilizes multi-band filtering and adjusts attention weights to improve the capture of temporal features.
• Spatial convolution extracts spatiotemporal feature sequences, optimizing temporal correlations through time dimensionality reduction.
• The model shows promise in adaptively extracting personalized features, potentially improving classification accuracy over existing methods.

Simplified