Preferential Subspace Identification (PSID) achieved the highest average correlation for predicting finger velocities (r = 0.43), outperforming XGBoost (r = 0.37) and SVR (r = 0.35).
PSID on high-gamma correlated channels allows for low-latency, accurate decoding of continuous finger speed from ECoG recordings.
Absolute Event Rate: 0.43% vs 0.37%
INTRODUCTION: High-density electrocorticography (ECoG) arrays have high spatio-temporal resolution for brain-computer interfaces (BCIs), but large data volumes introduce latency, preventing real-time decoding. For individuals with paralysis, restoring control of a dexterous external hand requires decoding strategies that are efficient and precise. METHODS: We collected high-density 1024-channel ECoG (Layer 7 Device, Precision Neuroscience) recordings over M1 (hand knob) from a single participant temporarily implanted as they transitioned between three hand gestures. Ground truth finger velocities were derived from a capacitive glove (StretchSense Pro Fidelity, StretchSense). The top-20 channels whose high-gamma band-power (100-200 Hz) had the highest Pearson correlation to finger kinematics were chosen. Using averaged power across 50-sample windows, ECoG data were split into frequency features: high-gamma, gamma (65-100 Hz), beta (11-30 Hz), and LMP (<3.5 Hz). Preferential Subspace Identification (PSID) extracted behaviorally relevant latent states to predict finger velocities. We compared PSID decoding performance to XGBoost, and Support Vector Regressors (SVR), which were trained on the same features. We evaluated performance using Pearson correlation of the predicted to the true signal. To reduce spurious predictions during rest, a Ridge Regression model trained on high-gamma and beta was used to gate predictions in all three models based on predicted movement onset. RESULTS: High-gamma correlation enabled channel reduction without sacrificing performance. Ridge regression gating improved model correlation by 0.03–0.06. PSID, using just 3 latent states, achieved the highest average correlation (r = 0.43), movement onset accuracy (a = 89%), and lowest BIC, outperforming XGBoost (r = 0.37, a = 89%) and SVR (r = 0.35, a = 81%). CONCLUSIONS: PSID on high-gamma correlated channels is an effective pipeline for extracting low-dimensional neural latent-states for decoding purposes, allowing for low-latency, accurate decoding of continuous speed.
Sargur et al. (2026) studied this question. Preferential Subspace Identification (PSID) vs. XGBoost and Support Vector Regressors (SVR) was evaluated on Pearson correlation of the predicted to the true finger velocity signal. Preferential Subspace Identification (PSID) achieved the highest average correlation for predicting finger velocities (r = 0.43), outperforming XGBoost (r = 0.37) and SVR (r = 0.35).