Neural Noise- BCI systems decode sensorimotor neural activity for continuous control of cursors or effectors; decoders perform well during intended movement but neural activity persists during intended non-movement, causing unintended activation and reduced stability.1
- Accurately identifying intended stationary vs movement states is a key requirement for stable BCI control.1 1
Weekly enrichment (2026-07-20)
- A University of Utah and Blackrock Neurotech team (Xiao, Kellis, Reiche, Solzbacher) published the cpSVM framework in Frontiers in Neuroscience on 20 February 2026; it gates a BCI decoder by classifying intended stationary versus movement states directly from intracortical multi-unit activity using PCA, correlation-based feature selection, and a linear-kernel support vector machine.2
- The offline study used a public dataset (DANDI 000070) from two rhesus macaques (J and N), each implanted with two 96-electrode arrays—one in dorsal premotor cortex and one in primary motor cortex, 192 channels total—performing a delayed center-out reaching task with 1,533 (J) and 1,530 (N) trials sampled at 1 kHz.2 3
- cpSVM reached mean classification accuracies of 0.936 (J) and 0.930 (N) under 10-fold cross-validation, roughly 8 percentage points above the conventional threshold-crossing baseline (0.866 and 0.859).2
- The classifier sharply improved output continuity, producing about 3 state transitions per trial (3.01 for J, 3.04 for N) versus 175 (J) and 211 (N) for threshold crossing, against roughly 2.3–2.7 transitions in the reference labels.2
- Because it reads neural activity that precedes movement, cpSVM detected state changes 16.5–88.7 ms earlier than the behavioral reference (predictive), whereas threshold crossing lagged behind; mean decode time was 0.647 ms per prediction.2
- The authors argue the gate belongs upstream of the decoder, since feeding rest-period neural activity into Kalman-filter or recurrent decoders can destabilize latent dynamics and the Kalman gain; stated limitations include offline-only testing, only two NHPs, and multi-unit-only signals.2
- The work builds on earlier idle-state detection: Velliste et al. (2014) showed that without explicit idle detection a prosthetic decode drifts about 200 mm/s during rest (twice reaching speed), cut to 3–4 mm/s once an LDA idle detector gated the output.4
- Complementary work by Williams et al. found local field potentials—especially beta and high-gamma bands—discriminate idle versus active intent on par with or better than firing rates, supporting pre-decoder state gating on Blackrock Utah-array recordings.5 6