DeepSMR: Decoding high-complex motor imagery via subject-dependent multi-feature refinement in deep convolutional networks

Electroencephalography (EEG) is a noninvasive neuroimaging technique that records electrical activity in the brain using electrodes placed on the scalp. It is widely used in neuroscience, clinical diagnosis, and brain–computer interface (BCI) applications to analyze brain signals in real time. This study proposes an advanced EEG-based BCI framework designed to decode and classify individual finger movements within a single hand during a finger-tapping task involving all five fingers. Our method employs a subject-dependent multi-feature refinement framework called DeepSMR, a novel deep convolutional network architecture optimized for feature extraction from EEG signals is introduced. This approach integrates spectral, temporal, and spatial analyses, leveraging event-related desynchronization/event-related synchronization (ERD/ERS), common spatial pattern (CSP), and power spectral density (PSD) techniques. Further, a subject-dependent multi-feature refinement framework. The DeepSMR achieved high classification accuracy for fine-motor tasks, achieving an average accuracy of 0.7471 (±0.0270) for the thumb and 0.7485 (±0.0314) for the index finger during motor execution tasks. DeepSMR outperformed EEGNet and DeepConvNet across all finger classes, showing an improvement of up to 15% in accuracy compared with the baseline models. Spectral feature analysis confirmed increased activity in the sensorimotor rhythm (SMR) frequency bands (8–13 Hz and 13–30 Hz), whereas temporal analysis revealed distinct patterns during the active and relaxed states. Spatial feature analysis highlighted class-specific features, further enhancing model performance. In the motor imagery session, DeepSMR maintained a superior performance, achieving the highest accuracy of 0.6984 (±0.0324) for the index finger. The results show that DeepSMR improves BCI performance by increasing the classification accuracy and computational efficiency, particularly for challenging finger-movement tasks. The framework could provide applications in neuroprosthetics, assistive robotics, and rehabilitation. In future work, the method could be expanded to include more motor tasks and integrate additional data types to further enhance the decoding accuracy for specific users and complex actions.