The adult bootstrap interval included zero for mean source prior gain after ten seconds of SEEG calibration in nine adults at a 500 ms horizon

We tested whether a model from another record helps after short SEEG calibration. We used 513 adjacent bipolar signals from nine adults in OpenNeuro ds003688 version 1.0.7. A fixed ridge model used 50 seconds from the other record and ten seconds of target calibration. The target-only model used the same calibration samples and penalty. Both models predicted identical later targets at a 500 ms horizon. The source prior raised mean skill by 0.0273 percentage points. The 95 percent adult bootstrap interval was -0.0154 to 0.0744 points, under independent adult sampling from the same population. The exact two-sided adult sign-flip p value was 0.3515625, under null sign symmetry. This test gives no clear support for a mean gain from this prior. It does not show equal performance. The transfer model had lower mean skill than the model with frozen source coefficients by 2.8403 points. This second comparison used 50 seconds of source training for both models. Record differences can include day and task changes. The selected adults and inputs were used in prior work. The package contains the compact inputs, native retrieval recipe, code, fits, checks, and results.

stereoelectroencephalography · SEEG · transfer learning · neural signal prediction · short calibration · linear regression · OpenNeuro ds003688