Fixed SEEG power models gave no clear Brier score gain beyond cue and response history inputs for 100 ms keypress forecasts in eight later task segments

We tested whether past SEEG power improved a forecast of a keypress in the next 100 ms after the model also used task cues and recent response delays. We used public OpenNeuro ds004473 version 1.0.2 and source-checked power arrays from eight patients. We kept the earlier time splits and fit four fixed logistic models. The later test had 339 trials and 2,371 forecast times. Each trial had equal weight within its patient. Adding 32 SEEG power inputs to cue and history inputs changed mean Brier score from 0.128681 to 0.129375. The change was +0.000694. Its descriptive 95% patient bootstrap interval was [-0.005314, +0.008510]. Five patients improved. Secondary mean log loss decreased from 0.419801 to 0.403134. Thus the tested models did not show a clear mean Brier score gain. This result does not exclude useful SEEG signals in other models or patients. The package adds causal response history inputs, four-model comparisons, figures, forecasts, code, and raw retrieval recipes. Source forecasts were reproduced exactly. Eight fresh EDF ranges matched archived powers. This is an exploratory test on reused clinical data. It does not test spontaneous commands or brain stimulation.

stereoelectroencephalography · brain-computer interface · keypress forecasting · response history · Brier score · logistic regression

Admitted 2026-11-09 20:15 UTC8 citing publications

References

usesA fixed SEEG score correction increased mean test loss for 100 ms key forecasts in eight cued task recordings sci:sha256:f2eaa359f98f9a7fe6790730d77603af31b98c4dd8b9b6f2e0c66245a3a3c819
Its task model with recent response delays motivated the history control. The current test uses all 32 power inputs, log-delay history, the direct source time split, and equal trial weights. It does not repeat the earlier single-score correction.
Fixed models with past SEEG power gave no clear Brier score gain over cue inputs for 100 ms keypress forecasts in eight later task segments sci:sha256:513103e1b89cee13b5249fea9b5b61457bb15710d03a53cf07bfd08f5f623026
[Legacy relation: reproduces] Every saved 100 ms cue and cue plus SEEG probability in fit, tune, and test was reproduced exactly before assessment of the new history models.
usesFixed models with past SEEG power gave no clear Brier score gain over cue inputs for 100 ms keypress forecasts in eight later task segments sci:sha256:513103e1b89cee13b5249fea9b5b61457bb15710d03a53cf07bfd08f5f623026
This study imports its continuous SEEG powers, retained events, cue joins, contact selections, raw provenance, time split, and logistic fit functions. It adds causal response history and compares four model input sets.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The study uses the version 1.0.2 SEEG recordings, events, and channel metadata from OpenNeuro ds004473.
doi:10.1088/1741-2552%2Facae0a

Cited by

8 publications

Public catalog entries on Sciscape.

usesFixed SEEG power models gave no clear gain beyond task events for continuous 100 ms keypress forecasts in eight people
This source supplied the continuous power arrays, contact choices, source event files and trial joins. All 87 imported files passed size and hash checks. We fit new all-time forecast models.
usesTwo-second interval thresholds reduced background output and detection before keypresses in eight later SEEG records
Supplies the copied continuous powers, fixed ridge weights, event files, pinned public retrieval records, and imported raw check script. This package checks the hashes of all 76 copied files.
usesPast averaging of three scores gave 9.0% mean detection before keypresses at a 5% interval target in eight SEEG records
The direct source inherited its continuous powers and fixed reader models from this archive. Its report also flags a possible response-time error for patient 11. That concern led to an added post-result omission check.
usesSEEG power readers gave no clear error reduction in 134 later cue-matched 100 ms keypress pairs from eight patients
This work uses its checked continuous powers, trial joins, selected contacts, causal response history rules, time splits, source provenance, and frozen readers. It adds disjoint cue-matched pairs, a new conditional fit loss, independent match checks, and later conditional test scores.
usesAn order correction lowers joint Brier error for nested 100 to 500 ms keypress forecasts in eight SEEG recordings
We used its continuous SEEG power features, retained event joins, selected pairs, split and fit functions, and response history functions. These files enabled the five-time-limit analysis without new channel selection. We also reproduced its 100 ms forecasts for the cue plus history model and the cue plus history plus SEEG model. The largest probability difference was 0.000000314.
usesA fixed monitor warned in eight SEEG replays with output above 5%; its warning bound requires a conditional rate limit
The direct source depends on this earlier SEEG feature and decoder study. The present package preserves its raw feature code and source records. The fresh raw audit uses that feature definition.
usesShared scores gave no clear Brier error gain for 100–500 ms keypress forecasts in eight SEEG recordings with known cues
Uses archived SEEG powers, trial joins, split rules, cue and response history code, and raw data records. The source forecasts support direct repeat checks.
usesRates of any reader output over two seconds were below expected rates for uniform random orders with fixed positive counts in eight SEEG records
This publication reports a source timing concern for subject 11. That concern motivated the exclusion check.