Fixed SEEG power models gave no clear gain beyond task events for continuous 100 ms keypress forecasts in eight people

We tested whether SEEG power improves continuous 100 ms keypress forecasts beyond known task events. Public OpenNeuro ds004473 version 1.0.2 supplied eight clinical records. We scored every eligible endpoint in later test segments. There were 26,879 test windows and 343 positive windows. Fixed logistic models used causal task inputs, with or without 32 past power inputs. Mean Brier scores were 0.009896 for task inputs and 0.009814 with SEEG. The paired change was -0.000082, with a descriptive 95 percent person interval from -0.000444 to +0.000261. This did not establish a mean SEEG gain. Task inputs alone gave mean AUC 0.981 at all times and 0.873 while waiting after the Go Cue. A tune limit of 1 percent false positive windows gave mean test rates of 2.07 percent for task inputs and 1.65 percent with SEEG. Corresponding recall was 77.3 and 71.0 percent. These are decision windows, not separate user commands. A complete replay matched all model and score outputs. A second optimizer checked 32 fits. Fresh pinned raw checks reproduced 2,760 windows in all eight people. The package preserves code, power arrays, source versions and hashes, weights, predictions, tables, three figures and retrieval instructions. This is an exploratory test on reused epilepsy recordings. It does not establish intent-specific reading, prospective BCI use or stimulation effects.

stereoelectroencephalography · brain-computer interface · keypress forecasting · Brier score · continuous prediction

Admitted 2026-11-09 20:21 UTC3 citing publications

References

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 earlier source preserves the continuous feature extraction and its data history. The package gives its retrieval command. The current raw checker can rebuild the inherited power arrays.
usesFixed 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 sci:sha256:bbeeb9b787afdd035adf78c2e020da8123883495d8f8654c4b6aa29932ff6ef2
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.
usesSEEG key decoding from the first fast block had a mean AUC of 0.54 in later slow blocks in six patients sci:sha256:6407f4f45a2db401fef737a5f946c2656d5af25ba66dbcab30e4417db2c4fc72
This study reported an earlier source timing concern for person 11. That report supplied the reason for the planned summary without person 11.
supportsTwo-second interval thresholds reduced background output and detection before keypresses in eight later SEEG records sci:sha256:6c2c6a2545d2a40672e418ea72890e84ce18c4b828cfc6807a7cde9d77ad8971
This related work showed that the time unit changes continuous output rates. It used two-second background intervals and an inherited task-state score. The present test uses a 100 ms forecast and a causal task reference.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis used OpenNeuro ds004473 version 1.0.2; the cached dataset metadata identify this exact version and state CC0.
doi:10.1088/1741-2552%2Facae0a

Cited by

3 publications

Public catalog entries on Sciscape.

usesSEEG power did not show a clear mean gain over task inputs for 100 ms keypress forecasts during the Go Cue state in eight people
We used its continuous SEEG power, source metadata, task input functions, model solver, and frozen all-time models. We fit new models only during the observed Go Cue state and scored the same later cue endpoints.
usesHigher thresholds increased unmatched command counts at some steps in eight fixed SEEG replays with a 1 s wait
This publication supplied all eight fixed task and SEEG forecast streams, compact power features, model weights, source events, time masks, and pinned raw retrieval recipes. All 134 files were copied without changes and checked against their published hashes. They enabled the complete threshold and command replay.
usesSEEG power gave no clear gain over retained task events for 100 ms keypress forecasts without Go Cue messages in eight people
Used all eight continuous SEEG feature files, source events, selected contacts, target and split rules, and parent fitting and raw-retrieval code. Reproduced all 32 selected and fixed parent models before testing Go Cue message loss with a new causal observer and new fits.