Publication

Abstract : Stroke and other neurological disorders often result in significant motor impairments, limiting the effectiveness of traditional rehabilitation. This project proposes a non-invasive Brain-Computer Interface (BCI) system designed to assist motor rehabilitation by decoding motor imagery (MI) tasks from EEG signals. Using the publicly available PhysioNet EEG Motor Imagery dataset, EEG signals from only three strategically selected electrodes (C3, Cz, C4) over the sensorimotor cortex are utilized, significantly reducing hardware complexity without compromising performance. Signals from 30 subjects are preprocessed to remove noise and artifacts, then transformed into 2D spectrograms via Short-Time Fourier Transform (STFT) to capture time-frequency features. A hybrid deep learning model is employed, where a Convolutional Neural Network (CNN) extracts spatial features from the spectrogram, and a Long Short-Term Memory (LSTM) network models the temporal dynamics for accurate classification of imagined movements (left,right and rest). The classified outputs are then translated into control commands and transmitted to an Arduino Uno microcontroller, which actuates a 4-DOF robotic arm using SG90 servo motors. This closed-loop system demonstrates a novel, low-cost, and adaptive BCI framework for hands-free robotic assistance and neurorehabilitation therapy. By leveraging an optimal number of subjects and selecting only three key motor cortex electrodes (C3, Cz, and C4), the system achieves effective motor intention decoding with reduced computational and hardware complexity.The deep learning model achieves a test accuracy of 85.89%, making it suitable for scalable and practical BCI deployment.