Review 25: Plug-and-play control of a brain-computer interface through neural map stabilization

Plug-and-play control of a brain-computer interface through neural map stabilization by Daniel Silversmith.

Silversmith, Daniel B., et al. “Plug-and-play control of a brain-computer interface through neural map stabilization.” Nature Biotechnology 39.3 (2021): 326-335.

  • Having recently read higher-level overviews of BCIs, reading this paper in Nature Biotech will be especially relevant but more specific.
  • Abstract
    • 180 channel electrocorticography (ECoG) implant in a paralyzed individual enabled stable monitoring of signals.
    • Long-term closed loop adaptation updates decoder weights over several days, thus updating map and plug and play control. Intro
    • Currently, 30 min daily recalibration results in variability in performance and decreased decoding performance in the long-term.
    • ECoG PNP control was achieved using an interface that updates within-session instead via closed-loop decoder adaptation (CLDA) and potentially over longer terms (ltCDA). Results
    • Fig. 1
      • ECoG signal comes from a few brain regions: S1, M1 , PMC, PMV, SMA.
      • Kalman Filter decoder weights.
      • There are a few aspects that are different between ltCLDA and daily initialization:
        • Time to taget increased, though it is hard to tell by how much as I can’t see mean values in the intervals.
        • ITR shows clear increase during ltCDA.
        • Convergence of decoder weights is faster using ltCDA.
      • Time to target decreases from 18 to 6 seconds with tight error bounds.
      • I do not understand box diagram in fig H. Are the outwward pointing arrows for CNS and KF pipeline outputs?
    • Weights are general orthogonal at the start of each dailty initializaiton, so pooling weights over days hurt model performance.
    • ltCDA resulted in stable convergence of decoder weights across sessions instead.
    • Higher half-life values (history length used in weight updates) led to better fixed control and more stable weight convergence.
    • Some weights exhibiting low magnitude or common trends may be redundant and enables sparseness.
    • Fig. 2
      • A lot more variance in decoder map weights for daily initialization method.
      • As the days go on, the orthogonality of the weights seems to decrease.
      • mu weights decrease w/ linear trends and gamma2 weights increase w/ linear trend.
      • Sparsity/weight pruning is pretty shocking with the 25% of the weights verison having a very similar time to target.
    • Fig. 3
      • Daily calibration had higher circular sd across channels in preferred direction (PD) task.
      • Decoder weight magnitude correlated to circular SD of PD.
      • gamma2 weights are stable from day 36+ and it is visually confirmed.
      • All weights show some trend toward stable convergence after day 18+, which is really interesting because it suggests long-term improvement with ltCLDA method.
    • Fig. 4
      • ltCDA to Fixed Decoder swap at day 206 + 28 recovery.
      • both ltCDA and CLDA after reset seem to have stable weights.
      • Decoder weights angle is roughly 45 degrees after reset. Fig. 5
      • Point and click task where cursor path and bits per second information rate are tracked.
      • Weights seem much sparser for click map.
      • I wonder how they will look for less convergent features.
      • Neural state needs to cross click threshold.
      • Everyday, after the cursor is in target area, stable clicking within 1 sec.
    • No significant drop in performance during PnP sessions (plug and play = no recalibration).
    • Decoder reinitialization had 45-degree weight angle change but then converge back to 0.
  • Discussion
    • ltCDA is an important example of weights being carried over a long time constant.
    • CLDA based on coadaptation can rapidly allow skilled control.
    • Task-related manifolds can be remarkably stable.
    • Results indicate the feasibility of PnP.