Vault

Closed-Loop Brain-Computer-Interfaces-for-Drug-Resistant-Epilepsy

3 min read

Dataview

LIST
FROM #Collection
WHERE file.name = this.Entry-For

Closed-Loop Brain-Computer Interfaces for Drug-Resistant Epilepsy: From Detection to Prediction and Control

🎤 Vocab

❗ Information

Small summary

✒️ -> Scratch Notes

Responsive NeuroStimulation (RNS)

75% median seizure reduction

  • what hapens to 25%? are there groups where symptoms worsen?

  • only 1/5 patients seizure free

  • months-years before reaching max efficacy

  • 1000 stim/day (3.5 years battery)

Instead of injecting energy, can we reduce enrgy?

Passive Neuromodulation (PNM)

  • circuit theory i dont understand
    • pass a transformation function to LFP in order to cancel out I
PNM for Temporal Lobe Epilepsy
  • Gate theory of denttate gyrus (DG)
    • Dentate region a target region for dealing with seizures
    • In mice:
      • Optogenetic inhibition stops forming seizures
      • Optegentic actiation starts seizures

Shows a model of DG activity

  • Weird divergent population activity (some super upregulated, others drop way down)
    • RNS doesnt change the trend
    • PNM does change the trend
  • Caveat:
    • PNM nead need to close to seizure onset zone (SOZ)
    • “proper SOZ targeting is important for single-site PNM”
  • Caveat can be minimized by:
    • Single to multi-site PNM
    • Increasing the number of contact pairs can relax the need for accurate SOZ localization
      • Do we activate all or inteligently?

?? What are behavioral effects of PNM ?

  • any logn term?
    ?? how does this affect surgery? is it more dangerous?
Robustness to delays in PNM onset
  • A lot more important for single site
  • Spatial and temporal robustness with PNM

“Fully” closing the loop with seizure detection

  • Timing and applied current

Epileptor

The epileptor - a widely used low dimensional neural mass model for simulating seizure-liek activity

  • 6 differential equations used to model seizure activity
    Inputs:
  • Something like current (I_stim)
    Output:
  • Something like LFP: LFP=x_1 - x_2

Open Qs:

  • Does this work experimentally?
    • electornics
    • full complexity and heterogeneity of epilepsies
    • soz targeting
    • electrode impedance, orientation, …
    • metal-electrolye barrier
    • safety
    • …
  • short and long-term plasticity
  • …

part 2:

reactive detection to seizure forecasting

seizure prediction:

  • NeuroVista Study (2013)
  • AES kaggle contest (2014)
  • nearly every ML classifier used since then

still about 80% accuracy today (even since contests)

  • not moving with pac eof machine learning
  • even these numbers highly inflated

their approach:

  1. iEEG window data
  2. extract features
  3. train on feature vector

use a feature dynamical model

Ceiling of prediction accuracy: role of scales

  • neural dynamics transorm massively across scales, so does their information conent
    (micro: neurons, meso: in-between (population/area-level/region), macro: brainscale (eeg/stereo-eeg))
    scale: micro -> meso -> macro
    complexity: monotonically decreases
    SNR: low -> high -> low

somewhere in between micro and macro.

  • claim: brain operates at meso scales

also challenging:

  • getting impedence low enough (sub 1 omhs?)

🔗 -> Links

Resources

  • Put useful links here

Connections

  • Link all related words

Graph View

Backlinks