🧠 BCI-Based Robotic Arm for Medical Rehabilitation

Published in IEEE | Amrita Vishwa Vidyapeetham, Coimbatore

---

📄 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, then transformed into 2D spectrograms via Short-Time Fourier Transform (STFT). A hybrid CNN-LSTM deep learning model classifies imagined movements (Left, Right, Rest), and the output is transmitted to an Arduino Uno to actuate a 4-DOF robotic arm using SG90 servo motors.

---

🏗️ System Architecture

EEG Dataset (PhysioNet)
        │
        ▼
Preprocessing (Notch Filter 60Hz + Bandpass 8–30Hz + Z-score Normalization)
        │
        ▼
STFT → Spectrogram (224×224×3 RGB — C3, Cz, C4 stacked)
        │
        ▼
CNN Feature Extractor (16→32→64 filters, BatchNorm, ReLU)
        │
        ▼
LSTM Temporal Classifier (256 hidden units)
        │
        ▼
Classification: Left | Right | Rest
        │
        ▼
Serial Communication (MATLAB → Arduino Uno)
        │
        ▼
4-DOF Robotic Arm (SG90 Servo Motors, PWM Pins D3,D5,D6,D9)

---

📁 Repository Structure

BCI-Robotic-Arm/
├── README.md
├── requirements.txt
├── src/
│   ├── preprocessing.py        # Notch + Bandpass filtering, normalization
│   ├── spectrogram.py          # STFT-based spectrogram generation
│   ├── cnn_model.py            # CNN spatial feature extractor
│   ├── lstm_model.py           # LSTM temporal classifier
│   ├── train.py                # Training pipeline (CNN 30 epochs, LSTM 70 epochs)
│   ├── predict.py              # Real-time inference
│   └── arduino_control.ino    # Arduino Uno servo control sketch
├── notebooks/
│   └── BCI_EEG_Pipeline.ipynb  # Full end-to-end notebook
├── results/
│   ├── confusion_matrix.png
│   ├── training_curve.png
│   └── ablation_study.csv
├── images/
│   └── system_overview.png
└── LICENSE

---

⚙️ Methods

Dataset

• Source: PhysioNet EEG Motor Imagery Dataset
• Subjects: 30 (first valid subjects)
• Electrodes: C3, Cz, C4 (sensorimotor cortex — ERD/ERS)
• Classes: T0 (Rest), T1 (Left), T2 (Right)
• Split: 70% Train | 20% Test | 10% Validation

Preprocessing

Step	Parameters
Notch Filter	60 Hz (powerline noise removal)
Bandpass Filter	8–30 Hz (mu & beta rhythms), 4th order Butterworth
Normalization	Z-score (zero mean, unit variance)
STFT	Hamming window=64, overlap=32, FFT=128
Spectrogram	224×224, RGB-stacked (C3, Cz, C4)

CNN Architecture

Layer	Details
Input	224×224×3 RGB spectrogram
Conv Block 1	16 filters, BatchNorm, ReLU, MaxPool
Conv Block 2	32 filters, BatchNorm, ReLU, MaxPool
Conv Block 3	64 filters, BatchNorm, ReLU, AvgPool
FC	128 units (spatial feature vector)

LSTM Architecture

Layer	Details
Input	128-dim CNN feature vector (as sequence)
LSTM	256 hidden units
Output	Softmax → 3 classes

Training

Stage	Epochs	Optimizer	LR
CNN	30	Adam	1e-4
LSTM	70	Adam	1e-4

---

📊 Results

Classification Performance

Class	Precision	Recall	F1-Score
T0 (Rest)	0.88	0.88	0.88
T1 (Left)	0.75	0.76	0.75
T2 (Right)	0.82	0.81	0.82

Model Accuracy

Model Variant	Accuracy (%)
CNN-LSTM (Proposed)	85.89
CNN Only	82.20
Shallow CNN	70.55
Shallow LSTM	82.10
No Batch Normalization	77.54

Comparison with Related Work

Model	Accuracy	Subjects	Electrodes
CNN-LSTM (This Work)	85.89%	30	3
CNN-LSTM (Fadel et al.)	70.64%	109	64
CNN+LSTM+DNN (Li et al.)	75.52%	12	64
CNN1D MF (Alnaanah et al.)	58.0%	105	64

Our model outperforms all baselines using only 3 electrodes — demonstrating hardware efficiency without sacrificing accuracy.

Training Time

• CNN Feature Extractor: ~4 minutes 56 seconds
• LSTM Classifier: ~35 seconds

---

🔧 Hardware Setup

Component	Specification
Microcontroller	Arduino Uno
Servo Motors	SG90 (4×) — Torque: 11 kg·cm at 6V
DOF	4 (lift, rotate, grasp, release)
PWM Pins	D3, D5, D6, D9
Communication	MATLAB → Arduino via Serial (UART)
Commands	LEFT, RIGHT, REST → servo angle PWM

---

🚀 Getting Started

Prerequisites

pip install -r requirements.txt

Run Preprocessing

python src/preprocessing.py --data_path ./data/physionet/

Train the Model

python src/train.py --epochs_cnn 30 --epochs_lstm 70 --lr 1e-4

Real-time Prediction

python src/predict.py --port COM3

---

📦 requirements.txt

See requirements.txt for full dependencies (numpy, scipy, torch, mne, pyserial, matplotlib).

---

---

📜 References

1. Goldberger et al. PhysioBank, PhysioToolkit, PhysioNet. Circulation, 2000.
2. Fadel et al. Multi-class classification of motor imagery EEG. IEEE BCI, 2020.
3. Li et al. Improving EEG-based MI classification using hybrid neural network. IEEE ICICN, 2021.
4. Pfurtscheller & Da Silva. Event-related EEG/MEG synchronization. Clinical Neurophysiology, 1999.

---

📍 Amrita Vishwa Vidyapeetham, Coimbatore, Tamil Nadu  |  Published in IEEE