In 8 SEEG subjects, decoder accuracy exceeded the best contact and tracked envelope-contact count, not total contacts

We compared archived results for eight subjects who passively listened to speech during SEEG. A cross-validated, high-gamma backward ridge decoder had higher speech-envelope correlation r than the best single contact in all 8 subjects (median difference 0.048; median ratio 1.19; exact sign-flip p = 0.0078). Decoder r correlated with the number of contacts whose in-sample r was at least 0.10 (Spearman rho = 0.93, exact p = 0.0024), the best contact r (rho = 0.95, p = 0.0011), and the fraction of contacts with r at least 0.10 (rho = 0.95, p = 0.0015). It did not correlate with total SEEG contact count (rho = 0.06, p = 0.90). This suggests that implant location and envelope-tracking coverage matter more than contact count for speech-envelope BCI, and that an in-sample single-contact screen may predict decoder accuracy. The conclusion is limited by eight subjects, correlated predictors, same-recording comparisons, a positively biased in-sample maximum for contact r, possible audio-to-neural timing error, and thresholds tested only at 0.10 and 0.20. Decoder r was cross-validated; part of its advantage may reflect its six temporal lags rather than more contacts. The analysis used archived result files and did not recompute features from raw recordings.

speech envelope · SEEG · high gamma · brain-computer interface · neural decoding · ridge regression · cross-validation · speech processing · electrode contacts