Full covariance gave no mean gain in offline four-site accuracy after stimulation contact order reversal with four SEEG pairs and five fit pulses per site in eight patients

We tested whether a fixed full noise covariance matrix improves reading of stimulation sites with a small SEEG calibration set. We used public OpenNeuro ds005448 data from eight patients. Each reader used four recording pairs, four known sites, and five fit pulses per site. Test pulses used the reversed stimulation contact order. Full covariance and a matched diagonal rule each made 117 correct decisions in 160 test pulses, or 73.1%. Their mean patient difference was zero. The descriptive patient bootstrap interval was -2.5 to 2.5 percentage points. Two patients improved, two became worse, and four had equal correct counts. This offline result does not establish equivalence or performance in a new session. Fit and test times interleaved across sites. We rebuilt all 320 source epochs from pinned public ranges. The package preserves features, models, code, three figures, source records, and checked reproduction instructions.

stereoelectroencephalography · SEEG · electrical stimulation · covariance shrinkage · site classification · noise covariance · offline benchmark