🧠 A818_NF — NeuroFlow

AI-powered brain signal decoder · Real-time EEG intent classification

Built by Asma Rizwan Rao

---

Overview

A818_NF (NeuroFlow) is a production-ready brain-computer interface software that decodes human intent from raw EEG signals. Dual-mode architecture supports both a high-accuracy 3-class decoder and an extended 5-class decoder with per-subject calibration.

---

Model Performance

Mode	Classes	Accuracy	Model	Calibration
Focus Mode	3 (Rest · Left Fist · Right Fist)	91.0%	Logistic Regression	Not required
Extended Mode	5 (Rest · L.Fist · R.Fist · Both Fists · Both Feet)	78.4%	EEGNet (CNN)	20 min per subject

• 3-class training: 2,820 trials, 5 subjects
• 5-class training: 4,500 trials (balanced), 20 subjects
• Data source: PhysioNet EEG Motor Movement/Imagery (109 subjects)
• Channels: 64 (3-class) / 22 (5-class, after harmonization)
• Sampling rate: 160 Hz

---

Architecture

Raw EEG Signal ↓ Preprocessing (MNE-Python) · Notch filter (50 Hz) · Bandpass (8–30 Hz) · Z-score normalization ↓ Feature Extraction / Deep Learning · 3-class: Flatten + StandardScaler · 5-class: EEGNet convolutional model ↓ Classification · Focus Mode → Logistic Regression · Extended Mode → EEGNet CNN ↓ Intent Prediction + Confidence Distribution

text

---

Software Features

• Dual-mode decoder — switch between 3-class and 5-class
• Two operating modes — Bench Test (saved data) vs Live Recording (hardware)
• Device-aware — Live Mode refuses to run without EEG headband connected
• Real-time EEG waveforms — scrolling plot for 6 channels
• Confidence distribution — per-class probability bars
• Session log — timestamped predictions with correct/incorrect flags
• Calibration workflow — per-class recording for personalized models
• Professional UI — dark theme, live status indicators, no fake outputs

---

Files

File	Purpose
app.py	Streamlit production app (dual-mode)
A818_NF_PhysioNet_v2.ipynb	5-class training notebook
NeuroFlow_v1.ipynb	3-class training notebook
best_model.pkl	3-class Logistic Regression model
scaler.pkl	3-class feature scaler
A818_NF_5class_eegnet_v2.pth	5-class EEGNet weights
A818_NF_5class_X_bal.npy	Balanced 5-class training data
A818_NF_5class_y_bal.npy	5-class labels
data/	Sample PhysioNet EEG files

Note: Large data files (.npy, .pth) are hosted on Google Drive due to GitHub's 100 MB file size limit.

---

Hardware Integration

The app includes a device interface layer for EEG headband integration. When hardware is connected:

1. Fill in EEGDeviceInterface.detect() — Bluetooth port scanner
2. Fill in EEGDeviceInterface.read_sample() — Bluetooth data reader
3. No other code changes needed

Recommended hardware: BioAmp EXG Pill + ESP32 (~$100)

---

Roadmap

• ✅ Phase 1 — 3-class decoder (91% accuracy)
• ✅ Phase 2 — 5-class extended decoder (78% with calibration)
• ✅ Phase 3 — Production dual-mode software
• 🔄 Phase 4 — EEG headband hardware integration
• ⏳ Phase 5 — Real-time live demo with users
• ⏳ Phase 6 — Medical partnerships & clinical validation

---

Tech Stack

AI/ML: PyTorch · scikit-learn · MNE-Python · NumPy · SciPy App: Streamlit · Matplotlib Training: Google Colab (T4 GPU) Deployment: Cloudflare Tunnel (dev) · Streamlit Cloud (planned)

---

Running Locally

pip install streamlit mne scikit-learn torch numpy matplotlib scipy
streamlit run app.py
Contact
GitHub: @deepbuildASI

LinkedIn: Asma Rizwan Rao

Email: asma.rizwanrao7@gmail.com

License
MIT License — see LICENSE