CompletedClassification2026-04-18

Neuromotion — motor imagery EEG (OpenBCI Cyton)

Open-source motor-imagery EEG stack for the OpenBCI Cyton (8 channels at 250 Hz, optional Cyton+Daisy at 16 ch / 125 Hz): BrainFlow acquisition, Graz-style recording with pygame, multi-session FIF pooling, pyRiemann training (tangent space + LR, MDM, CSP+LDA), and real-time classification with an inference UI. See the repository for CLI, SDK, montage, and signal-quality notes.

EEGBCImotor imageryOpenBCIpyRiemannBrainFlow

Code & media

View on GitHub

Live inference: motor-imagery three-class predictions. Here I was purely thinking about contracting my right hand, left hand or pressing my feet. I stated beforehand what I was going to think about to let the viewer know. Left and right work pretty well, it struggled to work with UP.

Summary

On real recordings for subject emi (three sessions, 290 epochs), stratified 5-fold cross-validation gave 44.5% ± 5.7% accuracy for three-class left / right / up motor imagery with a tangent-space + logistic regression pipeline (chance 33%). Binary left vs right with MDM reached 64.7% ± 6.7% (chance 50%) and is described in the project as usable for applications today. The weaker “up” (feet) class is expected on montages that do not emphasize medial sensorimotor coverage; the README discusses swapping cues, montage, and hygiene factors that dominate performance.

Key metrics

3-class acc (CV)

44.5% ± 5.7%

Tangent space + LR; chance 33%

Binary L/R acc (CV)

64.7% ± 6.7%

MDM; chance 50%

Epochs

290.0

Subject emi, 3 pooled sessions

Sampling

250 Hz

Cyton 8-channel default

Classes (3-cl.)

Left / right / up (feet imagery)

Code & data

GitHub

Acquisition through real-time inference

Experiment data excludes raw stimuli and large prediction arrays.

Back to experiments