MNE-RT: an open-source real-time neurofeedback/BCI framework

Comments

• wjcroft Mount Shasta, CA

  July 20 edited July 22

  Payam, wow.

  This is a fantastic contribution to the BCI / neurofeedback / EEG community. I see you are the primary author. To expand your links:

  Github:

  https://github.com/mne-rt-org/mne-rt

  Documentation:

  https://mne-rt-org.github.io/mne-rt/

  Modalities section:

  https://mne-rt-org.github.io/mne-rt/modalities.html

  Sensor Power
  Band Ratio
  ERD / ERS
  Laterality
  Laterality ERD/ERS
  Instantaneous Phase
  Slow Cortical Potentials
  Peak Alpha Frequency
  Hjorth Parameters
  Spectral Centroid
  Entropy
  Argmax Frequency
  Individual Peak Power
  Cross-Frequency Coupling (Sensor)
  Sensor Connectivity
  Connectivity Ratio
  Sensor Graph
  Source Power
  Source Connectivity
  Source Graph

  Academic Paper link:

  https://www.researchgate.net/publication/393419515_Advances_on_Real_Time_MEEG_Neural_Feature_Extraction

  Abstract:
  This paper introduces MNE-RT, a Python package designed for real-time neural feature extraction from magne-toencephalography (MEG) and electroencephalography (EEG) signals in Brain-Computer Interface (BCI) systems. The package incorporates efficient algorithms spanning traditional univariate metrics, such as frequency band power and entropy, to advanced bivariate connectivity measures. It is compatible with various recording systems, enabling the extraction of neural targets from brain signals in real time, with potential applications in enhancing neurofeedback efficacy.
  Published in: 2025 IEEE 38th International Symposium on Computer-Based Medical Systems (CBMS)
• wjcroft Mount Shasta, CA

  July 20

  A couple preliminary questions. With the 3D source localization features, do you think that the 16 channels of OpenBCI Cyton + Daisy would provide 'usable' resolution for a simple neurofeedback protocol that say rewards a threshold of amplitude at a given Brodmann / voxel area? I realize most 3D neurofeedback (sLoreta etc.) is done with 19 channels. How about the case with only 8 channels (seems unlikely).

  In a much simpler example neurofeedback program that say just rewards a threshold at a given scalp 10-20 location, what is the estimated delay time between the event at the EEG channel, and the triggered feedback? Just curious because in typical neurofeedback apps (BioEra, BrainBay, Bioexplorer, etc.) this usually a couple hundred milliseconds. Those apps are compiled code. Does the interpretive nature of Python, slow this down noticably?

  Regards, William
• Payam Switzerland

  July 22

  Hi William,
  Thank you so much for your interest.

  With the 3D source localization features, do you think that the 16 channels of OpenBCI Cyton + Daisy would provide 'usable' resolution for a simple neurofeedback protocol that say rewards a threshold of amplitude at a given Brodmann / voxel area?

  Technically, MNE-RT will build a forward/inverse model and run source localisation methods (e.g. sLoreta) on 16 channels, there's no hard channel-count gate. But of course it comes with resolution limits: with 16 sensors, the leadfield's effective rank is around 16, so the number of independent spatial patterns you can actually resolve is small, and the PSF (localization blur) for any given source is typically several centimeters which is not good for deep/medial sources (e.g. cingulate, insula) than for superficial cortex. So a literal single-voxel or narrow single-Brodmann-area threshold isn't really meaningful at this density becasue neighboring gyri will smear into each other in the estimate.
  That said, MNE-RT's NF modalities don't operate on raw voxels, they average activity within an anatomical label (atlas ROI, e.g. a Desikan-Killiany parcel), not a single dipole. That's actually the right granularity for low channel counts: a broad ROI (say, left vs. right sensorimotor cortex, or a whole frontal label) is far more robust to the blur than trying to isolate one small patch from its neighbors.
  For 16 chs configuration, my recommendation would be well-separated ROIs rather than fine-grained Brodmann-level specificity, you'll get plausible, stable lateralization/regional contrasts, not precision imaging.

  How about the case with only 8 channels (seems unlikely)

  At 8 channels, I'd honestly steer away from 3D localization claims entirely, you're down to maybe 2–4 very broad distinguishable regions at best, which is more "spatial filter" than "source image. Sensor-space or a single coarse bilateral contrast would better choice there.

  what is the estimated delay time between the event at the EEG channel, and the triggered feedback? Just curious because in typical neurofeedback apps (BioEra, BrainBay, Bioexplorer, etc.) this usually a couple hundred milliseconds. Those apps are compiled code. Does the interpretive nature of Python, slow this down noticably?

  For a simple 10-20 montage amplitude/threshold reward, the dominant term isn't Python's interpreter, it's the acquisition window itself. MNE-RT processes windowed chunks with 50% overlap (default 1 s windows, so a new estimate every 0.5 s), and that window length is a free parameter: for a plain amplitude threshold (no need for frequency-resolved band power), you can shrink it to e.g. 0.25–0.5 s without hurting stability, since you're not FFT-resolution-limited the way you would be for narrow-band power. The actual per-window computation runs through NumPy/SciPy (vectorized C/Fortran under the hood), typically sub-millisecond to a few ms per window on ordinary hardware. On top of that there's a small display-refresh cost (the live plot updates at up to ~30 fps) and whatever latency your acquisition hardware/driver itself adds before samples even reach LSL. For Cyton+Daisy specifically, the Daisy's interleaved dual-board sampling scheme adds its own reconstruction delay, and that's outside any software's control. I guess with a short window and a simple amplitude criterion, you should land in the same "couple hundred ms", window length and hardware buffering dominate, not language choice. I haven't yet checked hardware-validated end-to-end latency numbers (it's an open item on my roadmap). Please feel free to open an issue add any feature request here.
  Thanks again!
  Payam
• wjcroft Mount Shasta, CA

  July 22

  Payam,

  Thanks for these great replies.

  William