Shape scores showed no clear mean AUROC gain for uncalibrated sites at 20 to 100 ms in five offline 16-site SEEG tests

We tested whether removal of total waveform size improves detection of a response from an uncalibrated stimulation site. The source was OpenNeuro ds004457 version 1.0.2. Five reused clinical sessions supplied 183 eligible sites and 732 distinct test pulses. Each patient had 100 matched random libraries. The primary test used 16 known sites and 20 to 100 ms voltage. The shape reader had mean AUROC 0.592; voltage distance had 0.584. The difference was +0.008, with a descriptive 95% patient interval of -0.037 to +0.050. The data showed no clear mean detection gain. The shape cutoff accepted 85.3% of uncalibrated responses; voltage accepted 80.1%. Shape scores, labels, and acceptance stayed fixed when complete test vectors had gains of 0.5 and 2. A proof covers positive common gain for nonzero vectors in exact arithmetic. The archive preserves code, contact bins, event roles, all scores, cutoffs, source hashes, a raw retrieval recipe, and three figures. Fresh raw reconstruction and a complete replay matched all saved numeric results. The within-site split is not a global forward time split. This test does not establish a live BCI benefit.

stereoelectroencephalography · electrical stimulation · response recognition · uncalibrated sites · signed cosine · AUROC · common gain