Training rank maps gave no clear probability error gain in eight later-trial SEEG tests

We tested whether a fixed training rank map reduces probability error for a small SEEG reader. We used public OpenNeuro ds004473 version 1.0.2 data from eight people. Each reader used the same eight bipolar pairs and 24 power features. It learned from 1,332 earlier trials and was tested on 674 later trials. Mean Brier error was 0.19014 with standard scaling and 0.18914 with the rank map. The mean gain was 0.00100. Its descriptive 95 percent person-bootstrap interval was -0.00312 to +0.00542. The exact two-sided sign-flip reference p value was 0.6406 under independent people and a sign-symmetric null. Four people improved and four had higher error. The test gives no clear evidence of a mean benefit and does not show equal error. We reused a source cohort and did not test a live trigger, a separate day, or stimulation. A full public raw-range rebuild matched all 3,133,395 source feature values. A new process matched all numerical result files. The package keeps the fixed plan, source data and code, all maps and predictions, label controls, provenance, three figures, and instructions to repeat the work. This version corrects the file check after sci retrieval. Scientific results are unchanged.

stereoelectroencephalography · brain-computer interface · probability error · Brier score · rank transform · logistic regression · chronological test · OpenNeuro

Admitted 2026-11-09 17:10 UTC2 citing publications

References

supersedesTraining rank maps gave no clear probability error gain in eight later-trial SEEG tests sci:sha256:94f6a09513bb34a330be655a5baa388678bd22a046c6d39732f0fed5148c0064
This package corrects the manifest checker for the metadata folder added by sci retrieval and Python caches. All source data, models, predictions, scores, controls, and figures are unchanged.
usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
We used all eight source feature files, trial and quality rules, chronological split, eight-pair selection rule, standard logistic reader, and raw range extractor. The new test compares standard scaling, a training rank map, and fixed clipping. The standard reader reproduced the archived source probabilities. All raw features were rebuilt from pinned public ranges.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
This is the recorded DOI for the exact OpenNeuro ds004473 version 1.0.2 used to obtain the eight SEEG recordings and source feature data.
doi:10.1088/1741-2552%2Facae0a

Cited by

2 publications

Public catalog entries on Sciscape.

usesA 64-window rolling rank map raised mean Brier error in eight later SEEG tests
We used the fixed training-rank rule, fitting settings, pair choice, and later-trial split as the fixed reader. Our 1,348 fixed-map probabilities matched its archived rank-reader probabilities within 1.3e-13. We added causal rolling maps and controlled gain and state-frequency replays.
usesSame-trial power subtraction did not show a clear Brier error gain for eight SEEG pairs on later blocks in six recordings
Its Brier error comparison informed our main probability score. Its raw rebuild informed the full public range verification. We used new key labels and condition-specific block splits, with no copied rank maps or fitted readers.