Two calibration pulses gave lower mean four-condition accuracy than five at four SEEG pairs in eight STReEF patients after recorded contact order reversal

We tested the number of calibration inputs for a compact reader of brain stimulation responses. We used public STReEF SEEG data from eight patients, four fixed conditions per patient, and four selected bipolar recording pairs. Two early pulses per condition fit the pair selection, condition means, and noise scales. Five separate later pulses per condition tested the reader. The recorded stimulation contact order reversed between the groups. The features describe the voltage shape from 20 to 100 ms after each pulse. We compared this reader with the same method fit on five early pulses per condition. The two-pulse reader identified 110 of 160 test pulses correctly (68.75%). The five-pulse reader identified 122 (76.25%). The mean difference was -7.5 percentage points. Accuracy fell in five patients, stayed the same in two, and rose in one. The descriptive patient bootstrap interval was -16.25 to -0.625 points. The exact patient sign-flip test gave p = 0.15625. This small sample did not establish a population loss or equivalent performance. All ten two-pulse subsets gave aggregate scores below the five-pulse score. Mixed fits showed that extra data for pair selection also changed accuracy. These fits used five pulses and cannot serve as complete two-pulse methods. The package contains all features, predictions, fitted readers, figures, source hashes, and replay code. Fresh retrieval checked all 320 voltage epochs. The cohort and test pulses were reused from prior work. The test does not establish online performance, a neural mechanism, or the effect of actual pulse omission.

stereo-EEG · intracranial EEG · electrical stimulation · calibration · signal classification · brain-computer interface