Shrinkage reduced five-pulse waveform prediction loss in seven limbic SEEG patients after fixed source correction

We tested a fixed rule that uses variation among five training pulses to shrink their mean waveform toward zero when predicting five separate pulse responses. In 80 retained stimulation groups from seven patients in OpenNeuro ds004696 version 1.0.1, mean normalized loss in the 15–500 ms post-stimulation window decreased from 0.27181 to 0.21013, and all seven patients had lower loss. The mean gain was 0.06168, with a conditional 95% patient bootstrap interval of 0.04351–0.08174; the exact two-sided patient sign test gave p = 0.015625 under assumptions of independent patients and null sign symmetry. Shrinkage also beat zero and half-amplitude predictions in the primary window. The larger pre-pulse gain, including a pre-pulse loss of 0.22543 for shrinkage versus 0.20598 for zero, limits a stimulation-specific interpretation. Source correction used all exported trials before the split, so this evaluates waveform prediction after fixed correction rather than an independently preprocessed pipeline. The package retains the full reported methods, results, proof, exact trial inner products, data, and figure. GPL-licensed scripts are omitted to meet SCI code terms; the equations and cached inputs describe the numerical analysis, but executable reproduction code is not included.

SEEG · intracranial stimulation · waveform estimation · shrinkage · prediction loss · trial variability · OpenNeuro