Neural Noise- Stable neural states in motor cortex support robust handwriting decoding for BCI neuroprosthetics.1
- Long-term stability of handwriting-related neural representations is critical for chronic BCI use.1 1
Weekly enrichment (2026-07-20)
- The primary study (Nature Human Behaviour, 2025) used intracortical Utah-array recordings from human motor cortex while a participant attempted handwriting of Chinese characters (n = 306 characters, each averaging 6.3 ± 2.0 strokes).2
- The authors find motor-cortex activity evolves through a sequence of discrete, stable neural states, each corresponding to the writing of stroke fragments during complex handwriting.2
- Within a given state, individual neurons’ directional tuning curves stay stable, but their gain or preferred direction varies strongly across states — evidence that skilled movements are encoded through state-specific neural configurations.2
- Models that automatically infer the neural state and apply state-dependent directional tuning reconstructed recognizable handwriting trajectories with a 69% improvement over baseline models.2
- The decoder extends a velocity Kalman filter with a dynamic observation function: a pool of linear directional-tuning models is adaptively weighted and assembled based on Bayesian inference of the current state from incoming neural signals.2
- Versus a state-independent velocity Kalman filter, the state-dependent decoder cut root-mean-squared error by 13–18% and raised R² by over 69% (paired two-tailed t-tests, P < 0.001, large Cohen’s d), with consistent gains for single-unit and multi-unit activity and across letters, shapes, and numbers.2
- A related methods paper introduces StateMoE, a mixture-of-experts decoder with a recurrent router that models state transitions and reports a 6.92% accuracy improvement over prior methods on human intracortical handwriting recordings.3
- For context, an earlier intracortical handwriting BCI (Nature, 2021) let a participant with spinal-cord-injury paralysis type at 90 characters per minute with 94.1% raw online accuracy (>99% offline with autocorrect), motivating decoding of temporally complex movements.4