A stroke patient tries to move a hand. Nothing happens.
The brain sends the command. The hand does not answer, so the brain gets no sign that trying worked. strokesignal reads the attempt from the EEG and answers it.
795,000
people in the US have a stroke each year, and worldwide it is almost 12 million new strokes a year.
Sources: CDC, Stroke Facts (cdc.gov); World Stroke Organization, Global Stroke Fact Sheet 2025 (world-stroke.org).
The attempt is visible in the EEG
Imagining a hand movement quiets the rhythm over the motor area that runs that hand. The left hand is run from the right side of the brain (channel C4), the right hand from the left (C3).
A small model reads which hand
The model is standard: a band-pass filter at 8 to 30 Hz, common spatial patterns, then a linear classifier. It is trained on each session's calibration run and scored on a test run it never saw.
The answer arrives at once
When the model reads an attempt, the tendons draw and the fingers close, and a stimulation pulse is shown as the amber line along the tendons. The limb and the pulse are simulated: there is no patient and no stimulator behind this page.
Scroll back up and down to see the arm answer again.
It works on people who had a stroke
Accuracy on the test run when only the first k calibration trials are used.
Drive it yourself
Every trial below is a real recording. Pick a session and watch the model decide.
Press play.
-- 0 of 0 trials read correctly
What this does not show
- It is recorded, not live
- The EEG comes from g.tec's stroke-rehab recordings of three patients. Nothing on this page is a live brain.
- It is not a medical device
- The limb and the stimulation are simulated. No patient has used this, and it is not meant for treatment.
- It is not new
- g.tec's recoveriX already does brain-controlled stimulation for stroke rehab in clinics. This is a small, open, readable replay of the idea.
- Three patients prove little
- Accuracy changed by -6, +18 and +2 points between each patient's first and last session. That does not show that the therapy worked.
- Calibration is the cost
- Each session needs about 17 minutes of calibration recording before the model can be used.
- Real use cannot know the instruction
- In the recorded therapy, feedback was given only when the reading matched the instruction. A home system cannot know the instruction, so switch to "Real use" to see wrong guesses move the wrong hand.