Deep feature extraction and swarm-optimized enhanced extreme learning machine for motor imagery recognition in stroke patients

Background:

Interpretation of motor imagery (MI) in brain–computer interface (BCI) applications is largely driven by the use of electroencephalography (EEG) signals. However, precise classification in stroke patients remains challenging due to variability, non-stationarity, and abnormal EEG patterns.

New methods:

To address these challenges, an integrated architecture is proposed, combining multi-domain feature extraction with evolutionary optimization for enhanced EEG-based MI classification. The approach begins with subject-specific frequency band selection based on event-related desynchronization (ERD), aimed at reducing non-stationarity and improving signal relevance. Spatial and temporal features are then extracted using a combination of the scale-invariant feature transform (SIFT) and a one-dimensional convolutional neural network (1D CNN), providing a comprehensive representation of EEG signal dynamics. These features are fused and classified using an enhanced extreme learning machine (EELM), with hidden layer weights optimized using differential evolution (DE), particle swarm optimization (PSO), and dynamic multi-swarm PSO (DMS-PSO).

Results:

Experimental validation on a dataset of 50 stroke patients demonstrated an average classification accuracy of 97% using DMS-PSO with 10-fold cross-validation. Additional evaluation on the BCI Competition IV 1a dataset yielded 95% and 91.56% on IV 2a, indicating strong generalization performance.

Comparison with existing methods:

Unlike conventional BCI approaches, this method combines adaptive filtering, spatial–temporal hybrid feature learning, and metaheuristic optimization, resulting in a lightweight model with improved classification accuracy and robustness.

Conclusion:

These findings demonstrate the effectiveness of evolutionary optimization in dealing with the constraints provided by high-dimensional, non-stationary EEG data, making it a promising strategy for real-time MI classification in BCI-based stroke applications.