Donor correlations gave 0.005 lower mean AUC than local correlations with eight features and five fit pulses at 86 SEEG sites in eight patients

We tested whether background correlations from other patients improve a compact SEEG response reader after five local fit pulses. The source was OpenNeuro ds004696 version 1.0.1. We kept 86 selected sites in eight patients. Each donor fit excluded the target patient. Both readers used eight cosine features, the same local response template, and the same local noise scales. The first five pulses were fit data. The final five were test data. In the 20 to 500 ms window, equal-patient mean AUC was 0.9336 with donor correlations and 0.9387 with local correlations. The difference was -0.00513. A descriptive 95% resampling range was -0.01388 to 0.00369. All eight patient removal checks retained a negative mean after donor refitting. Donor correlations predicted later background correlations with less error in all eight patients. That gain did not improve primary mean AUC. Transfer effects changed sign across fixed feature and time checks. The evidence does not establish a population decrease or a live interface benefit. The package preserves voltage, source hashes, code, models, all scores, tables, three figures, and reproduction instructions. Fresh source extraction and an independent numerical calculation passed.

stimulation-evoked potentials · SEEG · response detection · AUC · covariance · background correlations · brain-computer interface