In 256 RAM SEEG patients, MNI kernel norms had low location skill and no plateau through 255 references

We measured how out-of-patient prediction by an MNI-location norm changed with the number of reference patients. The archived pair table contained 27,974 bipolar pairs from 256 RAM SEEG patients, with five spectral features from one 60-second task segment per pair. For each target patient, a Gaussian kernel norm (sigma 7.5 mm, 20 mm cutoff) was compared with the grand mean from the same reference patients. Mean R2 was below zero for all five features at 16 references and for four of five at 32. From 16 to 64 references, R2 increased by 0.134–0.200 (Holm-adjusted p = 0.00025 for each feature). R2 continued to increase from 128 to 255 references by 0.013–0.022; none met the prespecified plateau rule. At 255 references, mean R2 ranged from -0.000 for relative theta to 0.130 for relative beta. Adding a leave-one-pair-out patient offset raised R2 to 0.249–0.407, showing a larger patient-level component in these data. The asymptotic fit did not describe the curve well and was not used to extrapolate. These results are limited to one kernel without shrinkage, raw rather than patient-centered features, one task segment, and epilepsy patients with nonrandom implants. The reference subsets overlap at larger sizes.

stereo-EEG · SEEG · normative modeling · MNI coordinates · spectral features · kernel regression · learning curve · RAM study · patient-level variation