EEG-BASED BRAIN - COMPUTER INTERFACE FOR ASSISTIVE ROBOTIC CONTROL

Abstract

This paper presents a complete end-to-end software pipeline for classifying imagined motor movements from electroencephalography (EEG) signals without physical hardware. Using the BCI Competition IV Dataset 2a — 9 subjects, 22 EEG channels, 250 Hz, 4-class motor imagery (left hand, right hand, feet, tongue) — the system applies 8–30 Hz bandpass filtering, epoch extraction over a 0.5–3.5 s window, and Common Spatial Patterns (CSP) feature extraction via MNE-Python within a scikit-learn Pipeline. A session-wise evaluation protocol (train on Session T, test on Session E) yields an average cross-session accuracy of 38.46% across all 9 subjects, consistent with published CSP-based baselines for this benchmark. A Streamlit real-time simulation dashboard demonstrates live motor intent prediction with confidencegated stability control and CSV export.