MOTOR-IMAGERY EEG
This experiment classifies EEG recorded while people imagined moving their left or right fist, letting you inspect each prediction and see how confidence affects whether the model sends a cursor command
PUBLIC EEG · PHYSIONET · 160 Hz
01 / DECODING STUDIO
Run the model on a held-out recording to inspect its prediction and confidence
Test the same signal with a controlled disturbance
These repeatable disturbances are applied after preprocessing to test specific sensitivities, without simulating every real headset failure
Choose a trial and run the decoder
Predictions below this confidence level produce no command
A simulation using recorded EEG, with no headset or physical device connected
Each electrode is set to zero in turn to measure the change in the model’s probability of right-fist imagery
The schematic shows model sensitivity to channel removal, which cannot establish where a thought originates in the brain
Average power across the displayed channels, with the filter informed by motor-imagery research [1]; the spectrum alone cannot identify the imagined movement
02 / THE EVIDENCE
Participants remain separate across training, validation and testing so the evaluation measures generalisation to unfamiliar people
Participant-macro balanced accuracy
At the declared fallback confidence threshold
Balanced accuracy for this two-class task
| Model | Validation | Test |
|---|
Both columns report participant-macro balanced accuracy for every candidate, with the serving model selected using validation results only
Each point represents a different confidence threshold; high thresholds retain fewer trials, making the resulting accuracy less certain, so this exploratory chart should be interpreted alongside coverage
Each bar includes all eligible test trials for one participant, while the demonstration uses their first six valid trials without selecting for successful predictions
The selected model remains frozen while these disturbances measure sensitivity under specific conditions, without establishing suitability for a wearable or assistive device
Rows: recorded label · Columns: model prediction · All test trials before abstention
03 / RESEARCH
IntentLab investigates whether models can distinguish cued left and right fist imagery in previously unseen participants, using a fixed EEG window and allowing uncertain predictions to be rejected.
The comparison includes common spatial patterns [4] and a compact convolutional model inspired by EEGNet [5], with selection based on validation participants.
The selected model reaches 61.1% participant-macro balanced accuracy, although its confidence threshold fails the validation target for reliable control, as assessed in the research report.
Visual cues, eye movements and muscle artefacts may contribute to the predictions, so classification performance alone cannot demonstrate that the model has isolated motor intention.
A stronger evaluation would use repeated participant splits, an independent dataset and causal filtering for streaming signals, with a new untouched test set for each subsequent development.
The present results are restricted to a two-class, cued motor-imagery task and provide no evidence for decoding arbitrary thoughts, reconstructing dreams or measuring consciousness.
Read the methods, cited literature and critical assessment of the recorded results