A shape classifier gave 23 percent mean shaft accuracy after stimulation at separate stimulation contacts within runs from five SEEG patients

We tested a fixed waveform shape classifier in public SEEG from five patients. The source was OpenNeuro ds004457 version 1.0.2. Each test predicted the stimulated shaft within one patient run. Training excluded all pairs that shared either stimulated contact with the test pair. We removed the offset and positive voltage scale. Across 972 folds, equal patient mean accuracy was 23.043 percent after stimulation and 12.926 percent before stimulation. The uniform choice reference was 13.606 percent. The exact two-sided patient sign flip test gave p = 0.125 under patient independence and sign symmetry. This sample does not establish a population effect. An added descriptive comparison on common folds gave 33.557 percent after accuracy when shared stimulation contacts were allowed. The exact test pair remained absent. This comparison also increased training sample count. Fresh raw checks matched all 3,477,760 voltage bins. Independent calculations agreed on all 1,944 primary window predictions. A full repeat matched all twelve planned numerical files. The package preserves code, input windows, source versions, event identities, predictions, figures, and reproduction commands. The local plan had no external registration.

SEEG · electrical stimulation · waveform shape · shaft classification · transfer validation · patient-level inference