Enhanced EEG Classification of Foot Motor Imagery Using Unscented Kalman Filtering for Neurorehabilitation

Resumo

Electroencephalography-based brain-computer interfaces for lower-limb prosthetic control face significant challenges in motor imagery classification accuracy due to signal noise and artifacts. This study investigates the potencial use of Unscented Kalman Filter (UKF) to enhance foot motor imagery classification in virtual reality (VR) and non-VR environments. Six participants performed dorsiflexion and plantarflexion motor imagery while EEG signals were recorded using 16-channel OpenBCI. Data were processed with/without UKF preprocessing, then segmented into temporal windows (0.25s, 0.5s, 1s, 2s) with varying overlap. Feature extraction combined temporal (RMS, variance, Hjorth parameters) and spectral features (mu/beta band power and entropy). Three algorithms – Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and Knearest Neighbors (KNN) – were evaluated using leave-one-subject-out cross-validation. Results showed that UKF provided selective improvements, with significant accuracy gains for 1s overlapping windows (52-54% vs. 47-50%, p lt; 0.01) and SVM performance with 2s windows reaching 72.05% ± 8.71%. However, minimal differences occurred between VR and non-VR conditions (1-3%), challenging immersive environment assumptions. Individual variability was substantial, with Kalman filtering effects ranging from a 2.2% decrease to a 5.4% increase in accuracy, demonstrating participant-dependent filter effectiveness. These findings suggest preprocessing strategies should be individually optimized rather than universally applied, and simpler non-VR interfaces may suffice for effective motor imagery training.