The system uses a Filter Bank Common Spatial Pattern (FBCSP) feature extraction pipeline and compares three classifiers — support vector machine (SVM), k-nearest neighbors (KNN) and linear discriminant analysis (LDA) — on the BCI Competition IV Dataset 2a under a subject-dependent paradigm. With five-fold stratified cross-validation, SVM achieved the highest mean accuracy at 90.7% (±2.6%) and a Cohen’s Kappa of 0.876 (±0.035).
The classifier was then evaluated on an independent, unseen dataset through a replay-based hardware-in-the-loop (HIL) methodology, reaching 85.9% accuracy. A confidence-based command validation stage raised accuracy among accepted predictions to 97.6%, but those accepted predictions accounted for only 60.7% of commands.
The classifier output was interfaced with a physical differential-drive wheelchair prototype governed by an encoder-based closed-loop steering control scheme. To constrain the consequences of residual classification uncertainty, the system incorporates a multi-tiered safety strategy: confidence gating, self-terminating steering, transition braking, communication timeout supervision and a hardware emergency cutoff. The authors conclude that careful integration of classification, command validation and embedded safety design can yield a robust and practical EEG-based wheelchair control framework despite the imperfect reliability of motor imagery decoding.