A Deep-Learning Based Decoded EEG Neurofeedback Platform Using Muse-S

Open-source, portable neurofeedback platform using consumer-grade EEG and deep learning for real-time brain state decoding.

Project Overview

This project develops an open-source, deep learning-driven decoded EEG neurofeedback platform that transforms expensive, immobile fMRI-based neurofeedback into an accessible, portable solution using consumer-grade Muse-S headsets. The platform enables real-time brain state decoding and feedback delivery, making neurofeedback therapy more practical for clinical applications, especially for children and patients with movement difficulties.

Motivation & Impact

Traditional fMRI Limitations
  • $500-700 per hour cost
  • Requires 3-10 sessions across days
  • Immobile, scanner-bound
  • Unsuitable for frequent clinical use
Our EEG Solution
  • Negligible running costs
  • Portable & wireless (Muse-S)
  • No gel preparation needed
  • Real-time feedback capability

Technical Architecture

Platform Components

Data Acquisition

Muse-S headset with 4 dry electrodes (TP9/AF7/AF8/TP10) at 256Hz sampling rate

Real-time Streaming

Lab Streaming Layer (LSL) for synchronized EEG data and event markers

Deep Learning Decoder

EEGNet CNN architecture for real-time brain state classification

Dual-Branch System

Online Branch

Real-time neurofeedback delivery:

  • Live EEG streaming via LSL
  • 4-second sliding window analysis
  • Real-time decoder predictions
  • Visual feedback via PsychoPy UI
  • 6s induction → 2s fixation → 2s feedback cycle
Offline Branch

Decoder training pipeline:

  • Data collection & quality control
  • Event segmentation & labeling
  • EEGNet model training
  • Cross-session validation
  • Model freezing for online deployment

Model Performance

EEGNet Architecture Details

The decoder employs a compact convolutional neural network specifically designed for EEG-based BCIs:

  • Input: 2 channels (TP9, TP10) × 1024 time points (4 seconds at 256Hz)
  • Block 1: Temporal convolution (8 filters, 1×64 kernel) for frequency-specific features
  • Block 2: Depthwise spatial convolution (16 filters, 2×1 kernel) with ELU activation
  • Block 3: Separable convolution (16 filters, 1×16 kernel) combining temporal-spatial info
  • Output: Binary classification (motor vs. rest state)

Clinical Applications

Pediatric Therapy

Suitable for children with ADHD or autism who cannot tolerate gel-based EEG or fMRI scanners

Motor Rehabilitation

Portable neurofeedback for stroke patients and individuals with motor disabilities

Cognitive Training

Attention enhancement and working memory improvement through targeted neurofeedback

Mental Health

Potential applications in anxiety, phobia, and PTSD treatment protocols

Resources

Source Code
Research Poster

Update

I am currently working on the manuscript, and the compelte manuscript is expected to be done by the end of December 2025.