Three fit pulses per contact order gave 63.5% test accuracy for recorded stimulation blocks in eight SEEG patients

We tested a reader of responses to electrical stimulation in public SEEG data. The source was OpenNeuro ds005448, version 1.0.0. The fixed sample had eight patients and 96 conditions. Each reader used one recording pair and three fit pulses at each contact order. Four other pulses tested the reader. Mean test accuracy was 63.5% at 20 to 500 ms after stimulation. An independently selected reader before stimulation gave 51.0%. The mean gain was 12.5 percentage points. A descriptive patient bootstrap interval was 3.6 to 20.8 points. The exact patient sign-flip value was 0.0625 under the stated symmetry assumption. Six patients improved and two declined. A fixed 20 to 100 ms check gave 71.1% accuracy. A fit check after pair selection gave 91.1%, which was higher than the test score. Contact order changed with time block. Thus the data do not show a causal effect of contact order. This offline study does not test a live interface or useful brain writing. The package contains figures, fit and test tables, source data recipes, exact waveform inner products, and code. Native range hashes and independent calculations passed the checks.

stereo-electroencephalography · electrical stimulation · calibration · response classification · temporal confounding · brain-computer interface