Exploratory held-out two-component PCA checks gave 80.7% mean rejection of digitally doubled SEEG waveforms at 20–500 ms after five fit pulses in eight patients

We tested SEEG response checks with public OpenNeuro ds004696 v1.0.1 data from eight patients. The sample had 86 stimulation sites and 886 remote bipolar readout instances. Five early pulses fit each check. Five later pulses tested it at 20 to 500 ms. Rates give equal weight to patients and sites within each patient. PCA means principal component analysis. A frozen two-component fit accepted 8.8% of later responses; a symmetric common fit rank accepted 82.3%. This rank accepted 96.0% after simulated double waveform gain. An exploratory follow-up scored each of six responses against a fit to the other five. It accepted 81.8% of unchanged responses and rejected 80.7% after double gain, an added rejection of 62.5 percentage points (descriptive 95% patient interval 57.4 to 68.2). It accepted 86.2% after half gain. The follow-up used the original test cohort. Rank guarantees require exchangeability, which was not established. Numerical inputs, results, source hashes, and figures are retained. GPL analysis and extraction scripts are omitted, so the package cannot replay the calculations from the upstream links alone. The study does not test live BCI control or natural faults.

stereoelectroencephalography · principal component analysis · conformal prediction · residual rank · signal gain · brain-computer interfaces