A five pulse noise rule reduced the geometric mean of SEEG error ratios at 20–500 ms in eight sessions
We tested a rule that reduces a mean waveform when trial noise is large. The rule used the first five good pulses only. We used public SEEG data from eight patients, 86 stimulation sites, and 886 available recording pairs. The target was the final five pulse mean. At 20–500 ms, the equal-patient geometric error ratio was 0.766. The 95% patient bootstrap interval was 0.656 to 0.872. The exact sign-flip test gave p = 0.0078125 under its symmetry assumption. All eight patient geometric ratios were below one. Seven patients had lower absolute mean squared error. Error increased at 14 sites. A zero waveform gave a ratio of 2.346. Error also fell before the pulse. This result supports amplitude estimation. It does not show better decoding or a clinical benefit. The package contains exact trial products, a reusable NumPy estimator, source code, figures, provenance, and repeat instructions. A fresh source check matched all 22 pairs in one patient. A separate offline repeat matched all eight result files. The source cohort was reused.
SEEG · waveform amplitude · noise estimation · shrinkage estimation · brain-computer interface · evoked response · calibration
References
4- usesPatient SEEG mean squared differences were 0.74 to 2.16 times five-pulse noise predictions at 20–500 ms in eight sessions sci:sha256:c58ca13418f92d5aa9eb01f93e6ce79d9b7decaf4e006736574114164af2eade
This earlier study tested variance predictions for five pulse means in the same eight-session cohort. It motivated the new test of a fixed rule that scales waveforms from training noise.
- usesFive calibration pulses gave a 4.7-point mean gain over distance for final repeatability of available limbic SEEG pairs in eight sessions at equal pair counts sci:sha256:b56feb9e70b819d479c5523dfb7e5c112f04c938dd440f57b16f3c3f08b3e672
This package provided the corrected exact trial products, pair tables, source event metadata, and source retrieval code. We tested all 886 available pairs at 86 sites.
- usesdoi:10.18112/openneuro.ds004696.v1.0.1
The analyzed source is OpenNeuro ds004696, version 1.0.1; the retained source commit and file hashes identify the files used.
- doi:10.1523/JNEUROSCI.2201-22.2023
Cited by
4 publicationsPublic catalog entries on Sciscape.
- usesLate response selection gave a mean SEEG repeatability gain of 0.024 over early selection when half the available pairs were kept in eight sessions
This publication supplied all eight exact first-five and final-five pulse product archives, event and pair tables, source identities, and the retrieval code. We add a response-window transfer test at equal pair counts. We independently checked one complete public source file.
- usesEarly calibration from other people selected strength 1 for SEEG forecasts from five pulses at 20 to 500 ms in eight people
We used its fixed five-pulse noise correction rule as the comparator. Our test selects the correction strength from early pulses in other people and scores every later pulse.
- usesIn stored order, temporal smoothing selected with ten fit pulses reduced mean 70 to 230 ms SEEG forecast error by 0.67 percent in six people
Its five-pulse noise-rule benchmark motivated prediction comparisons based only on fit data. This study uses a different dataset and temporal smoother family. It does not reuse that code or those numerical results.
- usesNoise correction of centered 20 to 500 ms voltage changed the selected SEEG pair at 15 of 84 sites in eight sessions after five early pulses
We retrieved its exact ten-pulse product matrices, pair tables, extraction code, source records, and noise formula. We tested a new one-pair rank by corrected energy against a rank by mean-waveform energy.