Recent training data did not improve causal speech-envelope decoding in eight SEEG listeners

We tested whether training a speech-envelope decoder on passages recorded close in time to test passages improves decoding when the test passages and amount of training data are held constant. We used 16 passages from one 15–25 min passive-listening session in each of eight SEEG subjects from OpenNeuro ds004703. The causal decoder used common-average high-gamma amplitude (70–150 Hz) and 1–8 Hz voltage features, neural lags of 0–250 ms, channel PCA, and ridge regression. The planned primary contrast, near minus far training data averaged across forward and backward designs, was −0.001 in Pearson r (median −0.002; 4/8 subjects above zero; exact two-sided sign-flip p = 0.94). In the forward design, far training data about 11 min before test passages scored higher than near data about 4 min before in 6/8 subjects (mean difference −0.028; p = 0.055, not significant). Using 12 passages rather than 6 near passages improved forward-design decoding in 8/8 subjects (+0.038; p = 0.0078); averaged across designs, the increase was +0.022 (7/8; p = 0.0625). Twelve passages also outperformed 6 far passages in 7/8 subjects across both designs (+0.021; p = 0.039). In this single, short session, the amount of calibration data mattered more than its recency. These data provide no evidence that recent calibration data improve within-session decoding. The findings do not test drift over hours or days, and the analysis was offline in passive listening by epilepsy patients with clinical contacts.

speech-envelope decoding · SEEG · brain-computer interface · calibration · causal decoding · ridge regression · passive listening