Robust Motor Imagery-Brain-Computer Interface Classification in Signal Degradation: A Multi-Window Ensemble Approach.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) translate neural signals into commands, mimicking brain processing.
• Motor imagery (MI) BCIs decode imagined movements from sensorimotor cortex patterns.
• Signal degradation in mobile or low-resource settings limits current BCIs, especially at low sampling frequencies.

Purpose of the Study:

• To develop a robust MI classification framework resilient to signal degradation.
• To improve the practical applicability of BCIs in real-world scenarios.
• To enhance the biomimetic interaction of EEG-based BCIs.

Main Methods:

• Proposed a filter bank common spatial pattern with time segmentation (FBCSP-TS) framework.
• Classified four MI tasks (left hand, right hand, foot, tongue) using segmented EEG signals.
• Integrated spatial, spectral, and temporal dynamics with soft voting for segment-level predictions.

Main Results:

• FBCSP-TS outperformed CSP and FBCSP on BCI Competition IV datasets.
• Accuracy at 110 Hz was statistically similar to 250 Hz, demonstrating robustness.
• Optimal parameters (window=3.5s, move=0.5s) improved signal-to-noise ratio (SNR).
• External validation achieved 0.809 ± 0.092 accuracy and 0.619 ± 0.184 Cohen's kappa.

Conclusions:

• The FBCSP-TS framework effectively preserves MI-relevant neural patterns under degraded conditions.
• This advancement supports practical, biomimetic BCIs for wearable and real-world applications.
• The proposed method enhances resilience to noise and artifacts, crucial for mobile BCI use.