An eight-pair rule increased mean surviving high-gamma power after the worst simulated shaft loss in 36 SEEG test records
We tested selection of eight adjacent SEEG pairs in 36 patients from OpenNeuro ds004100 version 1.1.3. Ten seconds in one interictal record selected the pairs. A separate record supplied test power. The new rule maximized calibration power left after the worst complete loss of one shaft. Mean surviving high-gamma power increased from 40.3% to 45.6% of ideal intact target power. The paired mean gain was 5.28 percentage points. Its descriptive 95% patient bootstrap interval was 3.31 to 7.37 points. Intact power fell by 8.95 points. A simple cap of two pairs per shaft gained 3.96 surviving points and lost 5.50 intact points. We prove an exact cap method for nonnegative grouped scores. Integer sums preserve the objective and its tie rules for stored float64 weights. The package contains source features, native audit samples, frozen retrieval recipes, code, every patient score, loss scenarios, and figures. A second run matched all numerical files. Independent optimizers and fresh public voltage checks passed. This is an exploratory power benchmark in a reused clinical cohort. It does not show useful intent decoding, physical device reliability, or stimulation benefit.
stereoelectroencephalography · brain-computer interfaces · fault tolerance · feature selection · high-gamma power
References
4- usesWorst-loss training gave mixed later error changes compared with mean-loss training under known single-shaft loss in eight SEEG recordings sci:sha256:f96f0c9e819df690e154ea9350497b9f074430a21f2ac09a596a0f32153eadd3
Used the distinction between a better worst-loss calibration objective and later-record performance to define the test and its limits. No fitted model, code, or task data were reused.
- usesEight high-gamma SEEG pairs selected with ten seconds had 76.7% mean retention across two interictal records in 36 patients sci:sha256:4947f6220434c945bfb5f749fa56c65ce926ce8b7742925c1b030438e32ad73f
Used all 72 public SEEG feature files and the fixed 36-patient cohort. Preserved its full package, source metadata, digital audit samples, hashes, and voltage retrieval recipe. The new contribution is exact selection for one shaft loss and a separate-record surviving-power test.
- usesOne pair per shaft showed no clear gain in lowest balanced accuracy after a known simulated shaft loss in eight later-trial SEEG tests at 16 contacts sci:sha256:5f8c5a234d2daf29596b053b0c7a1a1f7aac5fdbdaf43faeac64da4b037739ad
Used the prior question of selection for a known single-shaft fault to define the grouped fault comparison and contact-budget limits. No code or data from this task study were used.
- usesdoi:10.18112/openneuro.ds004100.v1.1.3
This is the version-pinned HUP iEEG dataset used for all 72 SEEG feature files and the fixed 36-patient cohort. Its CC0 license and source provenance are preserved.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesSelection for the highest surviving score showed no clear Brier error decrease after known shaft loss in eight later-trial SEEG tests at 16 contacts
Adapt its total-minus-largest-shaft objective from surviving high-gamma power to early class scores. Add a constraint against shared contacts. Test whether the score gain transfers to later decoder probability error; do not infer decoded information from retained power.
- usesSelection for surviving task utility gave no clear gain in lowest later accuracy after known simulated shaft loss in eight SEEG recordings at 16 contacts
Its summed-power objective after the largest shaft loss supplied the selection idea. This study tested a task-effect surrogate with distinct contacts and measured later decoder accuracy. It did not reuse the interictal data or decoder evidence.