Averaging clipped calibration pulses gave no clear site accuracy gain in five SEEG sessions with a feature limit equal to fit RMS

We tested two calibration rules for stimulation-site decoding in public SEEG data from five patients. The software encoder clipped centered 5 ms bipolar voltage features. Its fixed pair limit was the fit RMS. One rule clipped the mean of four fit pulses. The other averaged the clipped fit pulses. Both used the same residual distance metric. The test used 552 later pulses from 138 sites on separate stimulation and recording shafts. In the 20–100 ms window, 64 site labels changed. Equal patient mean accuracy rose from 31.30% to 31.94%. The change was +0.65 percentage points, with a 95% patient bootstrap interval of -0.69 to +2.08 points and an exact sign-flip reference of p = 0.50. This did not show a clear mean gain. A secondary half-RMS limit favored the mean of clipped pulses in all five patients. A finite-sample identity shows why that mean minimizes encoded fit squared distance; it does not guarantee better test labels. All 110 pinned raw files passed hash checks. The raw rebuild matched all saved input arrays and event tables. A separate distance method matched all 33,120 method decisions. A separate process reproduced every numeric model and table. The package holds three figures, code, derived inputs, source metadata, fitted states, scores, and retrieval and reproduction commands. This is an exploratory offline test on reused epilepsy sessions. It does not test native amplifier clipping or a live brain-computer interface.

stereoelectroencephalography · SEEG · brain-computer interface · signal clipping · site decoding · calibration · intracranial electrophysiology