A distance-only CCEP response model did not transfer across two SEEG datasets without refitting

We fitted a logistic model of N1 response probability versus log distance on 18,857 contact rows from OpenNeuro ds005448 (8 patients, 1–2 mA, expert-checked N1, recording contacts outside white matter, 5–80 mm). Within ds005448, leave-one-patient-out tests showed calibration (observed-to-expected ratios 0.68–1.37, calibration slopes 0.63–1.41, AUC 0.77–0.94; fitted D50 13.8 mm, 95% CI 13.0–14.9). We applied the model without refitting to 2,573 rows from OpenNeuro ds004696 (4 patients, 6 mA, archived z-threshold N1 detector). At z > 3.4, the observed rate was 2.17 times the predicted rate, the calibration slope was 0.42 (0.24–0.68 per patient), AUC was 0.68, and the Brier score was worse than the null. Across higher detector thresholds, the observed-to-expected ratio changed from 2.17 to 0.26 and ds004696 D50 changed from 13.4 to 4.8 mm, while calibration slopes remained below 1. Common-median referencing gave the same pattern. The distance-response slope fitted in ds004696 was flatter (−1.25) than in ds005448 (−2.79), and contacts at 40–80 mm responded at 0.11–0.19 in ds004696 versus 0.01–0.03 in ds005448. These results show that this distance-only model did not transfer without local recalibration, and that cross-dataset spatial reach depends on the detection rule. The datasets differ in current, center, montage, channel sampling, tissue labels, and detector, so this comparison cannot identify the cause. The ds004696 detector had no expert check, only four test patients were available, and distance was Euclidean from the stimulation-pair midpoint.

cortico-cortical evoked potentials · stereo-EEG · calibration · distance model · N1 response · cross-dataset transfer · electrical stimulation