Toward Practical Wrist BCIs: Multi-Class EEG Classification of Actual and Imagined Movements

Brain-Computer Interfaces (BCIs) offer a promising solution for enabling the control of external systems. However, the traditional inter-limb approach is insufficient to achieve the dexterity required for daily tasks. The intra-limb approach, while more relevant for fine motor control, presents significant challenges in decoding fine intra-joint wrist movements, which are critical for dexterity, from non-invasive electroencephalography (EEG) signals. This study investigates the feasibility of classifying four wrist movements (flexion, extension, radial deviation, and ulnar deviation) from short EEG segments recorded during motor execution (ME) and motor imagery (MI). Data were collected from eight healthy participants using a 16-channel EEG setup. A lightweight machine learning pipeline, combining Common Spatial Pattern (CSP) and Filter Bank Common Spatial Pattern (FBCSP) for feature extraction, along with Support Vector Machine (SVM) and K-nearest neighbor (KNN) for classification, was validated using 10×10-fold cross-validation. The proposed pipeline achieved high performance in binary classification, with accuracies of 88.36% (ME) and 87.82% (MI) for flexion vs. extension. Comparable performance was observed for radial vs. ulnar deviation, with accuracies reaching 88.78% (ME) and 88.36% (MI). For the more complex four-class classification task, accuracies of 51.91% (ME) and 51.59% (MI) were obtained, significantly above chance level. Notably, MI performance was comparable to ME, likely due to the sequential ME–MI task design that enhanced neural priming. These findings demonstrate the feasibility of decoding fine intra-joint wrist kinematics from a lightweight EEG configuration, highlighting the potential for real-time, resource-efficient BCIs for neuroprosthetics, rehabilitation, and assistive technologies requiring fine motor control.