Adding power after a keypress lowered mean log loss in seven SEEG records with eight pairs per model

This test measured the added value of 70–110 Hz SEEG power after a keypress. The reference model could use task timing and power before the keypress. Each fit used eight pairs, with choice and training restricted to earlier blocks. Seven records supplied 979 later trials and 20 wrong keys. Mean log loss, with equal weight per record, fell from 0.108484 to 0.100841 nats per trial. One record supplied 83% of the reduction. At probability 0.05, the added model gave 7 correct alerts, 106 false alerts, and 13 misses. Its Brier score was worse, and its mean loss did not beat the past error-rate model. Subtracting earlier power left little loss gain. These reused task data do not establish a reliable signal for automatic BCI correction. The 36 MB package preserves code, features, source versions and hashes, fresh checks of public raw data, model fits, trial scores, figures, and reproduction steps. A second optimizer and a separate spectrum calculation passed numerical checks.

stereoelectroencephalography · brain-computer interface · error prediction · high-frequency power · logistic regression · log loss · false alerts · exploratory analysis