Eight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings

1. usesdoi:10.18112/openneuro.ds004473.v1.0.2

   This is the exact OpenNeuro ds004473 version used for the eight SEEG recordings and task metadata.

2. doi:10.1088/1741-2552%2Facae0a

Public catalog entries on Sciscape.

• usesBand power ratios retain 67.0% balanced accuracy under simulated constant voltage gains per shaft in eight SEEG recordings

  We use the preserved full eight-patient SEEG feature set, source metadata, pinned raw audit ranges, and MIT extraction code. We add fixed gain tests, two centered power models, and exact threshold bounds. This is a new analysis of the same sample.

• usesAdding a band contrast reduced mean test error beyond features outside the band in eight SEEG records

  Used its eight fixed bipolar pairs per person, retained trials, state windows, and early-versus-later split. The report states the limits from power-based pair choice and source quality selection.

• usesZero-centered 8-bit waveform codes reduced frozen SEEG task-state accuracy from 70.7% to 50.6% in eight recordings

  We used the frozen eight-pair models, trial rules, original features, metadata, source code, and pinned EDF byte-range recipes. Native waveforms reproduced the selected parent power features and test probabilities.

• usesSeparate beta calibration increased mean AUC in later trials from eight SEEG recordings with eight pairs

  This study uses the inherited physical bipolar extraction code, five-window spectral features, and chronological early-calibration split from this method.

• usesA shared SEEG model gave 52.6% mean balanced accuracy in eight tests that held out one person after rest calibration

  Used all eight preserved SEEG feature arrays, trial metadata, source records, exclusions, and raw retrieval code. Added a shared model that holds out each person’s task data. Reproduced the parent eight-pair balanced accuracies as a reference.

• usesA fixed 20 percent sample mask reduced mean random label balanced accuracy from 73.5 to 50.0 percent in eight SEEG rest recordings with simulated gaps

  This work made the original eight-pair selection, source state windows, trial rules, spectral features, and fixed logistic task models. We kept those inputs and reproduced native features and probabilities before testing the correction.

• usesSimulated one-sample contact delays changed a mean 6.89% of later decisions in eight fixed bipolar SEEG readers

  Used the trial set, chronological split, candidate features, selected pairs, fixed model weights, pinned source ranges, and MIT source reader.

• usesLinear interpolation gave higher mean balanced accuracy than pair-weighted power in eight fixed SEEG decoders at 20% expected loss of 10 ms packets

  This work used the eight early-trained SEEG state decoders first constructed in this publication. It preserved their pair selection, training split, scaler, coefficients, and threshold. The clean later-trial mean accuracy was reproduced before the new packet-loss test.

• usesLinked band bounds proved more SEEG classes stable under 5% constant digital contact gain errors in eight records

  The inherited eight-record benchmark supplies the early pair selection and rest versus pre-response feature construction. These features and splits were obtained through the spectral moment study and remain fixed here.

• usesFourfold filtered sampling reduced mean balanced accuracy by 0.58 points in eight offline SEEG decoders

  This publication supplies retained trials, early training features, pair selection, fixed scaler and decoder states, and the original feature definition. We keep these inputs fixed to test later-trial sampling changes.

• usesA paired contrast fit did not show a mean task-state AUC gain over ordinary ridge in eight later-trial SEEG tests

  Used all eight unchanged feature arrays, their trial and exclusion records, 94 source files, and two source retrieval scripts. All 118 imported files matched the parent hashes. They enabled the paired contrast versus ordinary ridge comparison and fresh raw-source audit.

• usesWorst-loss training gave mixed later error changes compared with mean-loss training under known single-shaft loss in eight SEEG recordings

  The study uses all eight exact feature files, trial records, exclusions, raw byte receipts, and preserved extraction and validation code. They provide the inputs and source checks for the later-trial fault comparison.

• usesBand power ratios retain 67.0% balanced accuracy under simulated constant voltage gains per shaft in eight SEEG recordings

  We use the preserved full eight-patient SEEG feature set, source metadata, pinned raw audit ranges, and MIT extraction code. We add fixed gain tests, two centered power models, and exact threshold bounds. This is a new analysis of the same sample.

• usesAn exact contact rule gave higher training scores and higher mean test Brier error in eight SEEG recordings at ten contacts

  We used the complete spectral features, trial records, source metadata, raw audit bytes, and retrieval code. The new test changes contact selection and model penalty choice. It keeps the source trial split and exclusions.

• usesKnown response times gave mean test gains below 0.02% of baseline loss in linear SEEG power models for eight people

  Its feature extraction code supplied the unchanged EDF parsing, bipolar pair, and byte-range retrieval utilities preserved in source_utils.py.

First page of citations