This paper examines how recognition of motor imagery (MI) in stroke survivors can be improved on the basis of EEG signals, which are noisy and unstable after stroke. The authors created a pipeline that first filtered the EEG by subject-specific band selection from event-related desynchronization. They then derived spatial features (using scale-invariant feature transform, or SIFT) and temporal features (using a one-dimensional convolutional neural network), merged them, and fed the results into an enhanced extreme learning machine (EELM). To improve performance, EELM was optimized using swarm intelligence algorithms like differential evolution, particle swarm optimization, and dynamic multi-swarm PSO. On a 50-patient stroke sample, the dynamic multi-swarm PSO variant recorded an average accuracy of 97% in 10-fold cross-validation. The model generalized to publicly available datasets nicely, too, at 95% accuracy on BCI Competition IV-1a and 91.6% on IV-2a. The authors believe that this metaheuristic optimization-based hybrid technique which combines adaptive filtering and deep learning, is a lightweight yet effective tool for MI classification in patients with stroke.