NeuroStream

Real-time BCI motor imagery streaming

NeuroStream trains a cross-subject EEG motor imagery classifier on the Zhou2016 dataset via MOABB, then replays a held-out subject as a pseudo-online stream. All feature extraction and classification pipelines are directly sourced from MOABB and pyriemann.

Highlights

• Four selectable pipelines via MOABB / pyriemann: CSP, FBCSP, TS+LDA, TS+SVM
• Euclidean Alignment (EA) for cross-subject covariance shift reduction (CSP and FBCSP)
• Riemannian geometry (pyriemann): covariance estimation → Tangent Space classifier
• Desktop UI built with tkinter, background training thread, root.after() driven streaming
• Uses an existing local MNE-zhou-2016/ folder when available — no forced redownload
• All parameters exposed in the UI and stored in BCIConfig

What The App Shows

Area	Purpose
Data Folder	Point the app at a local Zhou2016 download
Feature Extraction	Choose CSP / FBCSP / TS (Riemannian)
Classifier	Choose LDA / SVM
Spatial Filters	Set CSP filter count; set custom frequency bands for FBCSP
Train & Load	Build and fit the model without freezing the GUI
Live Feed	Pseudo-online trial countdown and prediction state
Confidence Bar	LEFT vs RIGHT class probability
Band Power	Relative mu and beta power for the current trial
Trial History	Recent hits / misses and cumulative accuracy line
Confusion Matrix	Running class-level performance (LEFT vs RIGHT)

Pipelines

All pipelines are implemented using MOABB and pyriemann building blocks.

CSP

Raw EEG
  -> Bandpass filter (configurable, default 8–30 Hz)
  -> Euclidean Alignment (He & Wu 2020)
  -> mne.decoding.CSP  (log-variance features)
  -> LDA or linear SVM

Classic MOABB baseline — Jayaram & Barachant (2018).

FBCSP

Raw EEG
  -> moabb.paradigms.FilterBankLeftRightImagery
     (configurable bands, default: 8–12, 12–16, 16–20, 20–24, 24–28, 28–32 Hz)
  -> Euclidean Alignment per band
  -> moabb.pipelines.utils.FilterBank(CSP)  — CSP applied per band, features concatenated
  -> LDA or linear SVM

Ang et al. (2012).

TS+LDA / TS+SVM

Raw EEG
  -> Bandpass filter
  -> pyriemann.estimation.Covariances (OAS estimator)
  -> pyriemann.tangentspace.TangentSpace (Riemannian metric)
  -> LDA or linear SVM

Barachant et al. (2013). Top-performing family in MOABB motor imagery benchmarks.

Quick Start

Requirements

• Python 3.9+
• macOS with tkinter available

Run (recommended)

bash run.sh

run.sh automatically creates a virtual environment, installs all dependencies, and launches the app. Safe to run multiple times — it skips setup if the environment already exists.

Manual setup (alternative)

pip install -r requirements.txt
python main.py

Dataset Layout

/your/data/path/
└── MNE-zhou-2016/
    ├── sub-1/
    ├── sub-2/
    ├── sub-3/
    └── sub-4/

If not present, MOABB can download it automatically on first run.

Typical Workflow

1. Paste or browse to the folder containing MNE-zhou-2016/.
2. Select Feature Extraction method and Classifier.
3. Adjust Bandpass / Epoch Window / Spatial Filters as needed.
4. Click Train & Load.
5. Click Start Stream to replay the held-out subject trial by trial.
6. Use Pause / Stop while inspecting live metrics.

Project Structure

APP/
├── main.py            Entry point (window init + mainloop)
├── config.py          BCIConfig dataclass — all tunable parameters
├── data_engine.py     MOABB data loading, paradigm selection, EA alignment
├── model.py           Pipeline factory (CSP / FBCSP / TS+LDA / TS+SVM)
├── streaming.py       Trial-by-trial streaming simulator
├── ui/
│   ├── app_view.py    Tkinter layout, state machine, training/streaming control
│   └── plots.py       Canvas drawing functions (confidence, band power, confusion matrix, chart)
└── requirements.txt

Default Parameters

Parameter	Default
Feature Extraction	CSP
Classifier	LDA
Bandpass	8.0 – 30.0 Hz
Epoch Window	0.0 – 3.0 s
CSP Filters	8
FBCSP Bands	8-12, 12-16, 16-20, 20-24, 24-28, 28-32 Hz
Train Subjects	1, 2
Test Subject	3

References

• Zhou2016 dataset: Zhou et al. (2016). A Fully Automated Trial Selection Method for Optimization of Motor Imagery Based BCI. PLOS ONE.
• MOABB: Jayaram & Barachant (2018). MOABB: trustworthy algorithm benchmarking for BCIs. J. Neural Eng.
• Euclidean Alignment: He & Wu (2020). Transfer learning for BCIs: A Euclidean space data alignment approach. IEEE Trans. Biomed. Eng.
• FBCSP: Ang et al. (2012). Filter bank common spatial pattern algorithm on BCI competition IV. Front. Hum. Neurosci.
• Tangent Space / Riemannian geometry: Barachant et al. (2013). Classification of covariance matrices using a Riemannian-based kernel. Signal Processing.
• CSP: Blankertz et al. (2008). Optimizing spatial filters for robust EEG single-trial analysis. IEEE Signal Process. Mag.

License

MIT. See LICENSE.