FROM SIGNAL TO A MEASURED RESULT

Your first decoding experiment.

Try real EEG, compare methods, and understand what the results can tell you. No setup or account needed for the example.

01 Choose data
02 Check & run
03 Understand
04 Take it forward

01 / YOUR DATA

A clear task, to start.

Left or right fist imagery

Predict the movement a participant was instructed to imagine.

Example dataset

Loading the prepared example…

3 electrodes: C3, Cz, C4 · 128 samples/second · 3-second windows

Bring your own data

Prepare labeled, de-identified epochs and train in Python. Import the exported experiment.json to run it locally.

Use C3/Cz/C4, 128 Hz, 4–38 Hz filtering and 384 samples per epoch.
Keep recording groups separate. Fit preprocessing only on training data.
Use the supplied exporter; retain an untouched test set.

Raw EDF and CSV uploads are not supported. Files with compatible shapes still need compatible tasks, labels and preprocessing.

Start-here notebook ↓Python training & export script ↓Preparation instructions ↗

Maximum 12 MB. Model and EEG stay in this tab.

02 / CHECK & RUN

Check the starting point.

Data checks will appear here.

What these checks mean

We check structure, known labels, class balance, flat channels and unusually large normalized samples. These are preparation checks, not a complete artifact review.

Normalized samples cannot establish electrode contact, physical voltage, correct labels, or an independent split.

Decoding method

Spatial filters plus a simple linear classifier. Recommended using validation recordings, before the final test.

Computation runs in this browser. Training is a separate, reproducible Python workflow.

How is the test kept separate?

For each participant, run 4 trains and run 8 validates the method choice. The selected method is refitted on runs 4+8, then tested on run 12. The default is frozen before run 12 is evaluated.

These are later runs from the same participants. This does not test new people, new days, or clinical performance.

Evaluation record & limitations ↗

03 / UNDERSTAND THE RESULT

Let the data answer.

COMPUTED LOCALLY

Ready when you are.

One click. A real comparison.

Run the example to get measured predictions, a simple baseline and a plain-language explanation of the limits.

The illustration above is decorative, not measured EEG.

YOUR INTERPRETATION ASSISTANT

Make sense of the next step.

EXPLAINS, NEVER DECODES

Ask about a term, the setup, or the results you just calculated. Suggestions are not measured findings.

Assistant modeChecking availability…

Hosted AI sends your question, up to four recent messages and selected numerical result metrics to OpenAI. EEG and model files stay local.

Privacy, limits & optional local AI

Hosted AI: 3 questions/minute and 20/day per browser, with shared limits of 30/minute and a $20 monthly application allowance. Limits or connection errors switch to labeled guided help. Responses are capped. No EEG uploads or chat text are stored in the site's accounting ledger.

Guided help uses curated explanations with no generation. A local reply was generated on one tested device. Local AI remains experimental and its answer quality is unvalidated.

Full privacy information ↗

04 / TAKE IT FORWARD

Leave with something useful.

Export the actual measurements and a record of the method, split and limitations. No raw EEG is included in result exports.

Open standalone in a new tab for downloads ↗

Inspect

Look at class errors and variation across people. A single average can hide failures.

Freeze

Choose methods on validation data. Keep your next test untouched until the choices are fixed.

Reproduce

Use the source workflow with your own task, recording groups and documented permissions.

Download the full source & Python workflow ↓