In 28 SEEG patients, refitting the intercept on two stimulation sites lowered held-out log loss, but half the gap to a patient model remained

We evaluated whether calibrating a cross-dataset distance-only model with a small number of stimulation sites improves prediction of N1 responses at other sites in a new patient. We used archived site-by-contact tables from three OpenNeuro SEEG datasets (28 patients total): ds005448 (8 patients), ds004696 (4), and ds004080 (16). For each target patient, a model of logit P(N1) = b0 + b1 × log(distance in mm) was fitted on the other two datasets, then tested on the target patient after calibrating on k = 1, 2, or 4 randomly selected sites (50 draws per patient). In the primary, prespecified contrast at k = 2, fixing the source slope and refitting only the intercept reduced mean held-out log loss from 0.643 to 0.578; 15 of 28 patients improved (two-sided sign-flip p = 0.015). This closed a median 50% of the gap to a patient model fitted on all other sites (log loss 0.498). At k = 1, the gain over no refit was not significant (p = 0.14); at k = 4, intercept-only and full refits performed similarly (p = 0.91). The refit worsened predictions for ds004696 patients (0.481 to 0.505), did not correct between-dataset slope differences, and left under-prediction at held-out sites (median observed/expected ratio 1.37). Dataset N1 definitions and stimulation currents differ, so the results cannot separate physiological variation from label-definition effects. The findings support a partial improvement from two to four calibration sites, with the limitations of small samples, random site selection, and a distance-only model.

cortico-cortical evoked potentials · SEEG · N1 response · electrical brain stimulation · model calibration · logistic regression · cross-dataset transfer · held-out log loss