Maximum score calibration left a mean absolute rate error of 8.2 percentage points at a 10% combined target in eight SEEG session pairs with 32 contact pairs each
We tested whether a new maximum score threshold reduced later combined call rate error beyond a new median for each contact pair. We used public OpenNeuro ds005624 version 1.0.0. The test retained all eight people with two sessions and 32 fixed contact pairs each. Separate 60-second blocks fit centers and thresholds. Each second-session test used 300 later one-second windows. At a 10% combined target, the new maximum rule gave a mean rate of 15.83% and a mean absolute rate error of 8.17 percentage points. New medians with the source threshold gave 11.25 points of error. The paired benefit was +3.08 points. Its descriptive 95% patient-bootstrap interval was -6.46 to +13.25 points. The exact sign-flip p value was 0.6484. This does not establish a benefit across people. Separate 10% thresholds for each pair gave an 89.04% combined rate; that rule has a different target. Under conditional exchangeability, the maximum rank rule has a next-call bound of 6/61. The records do not establish that assumption. There are no task labels or stimulation responses, and voltage scale is unresolved. Calls describe exported numbers, not intention detection. The package preserves all nine fixed settings, source arrays, raw witnesses, code, hashes, fitted models, decisions, and four figures. All 576 pinned public byte ranges were checked and all seven source features were rebuilt exactly. An independent scalar replay checked all 86,400 decisions. A clean replay gave identical numerical output bytes.
stereoelectroencephalography · SEEG · brain-computer interface · high gamma power · split conformal calibration · call rate · signal processing · OpenNeuro
References
3- usesA 120-second median calibration left a mean high gamma rate error of 5.1 percentage points for a 10% target in eight SEEG session pairs sci:sha256:878731c694843c59980d6304a735b1a04b5ce68504f08fef057fdf9c73485346
This test uses its eight feature archives, 16 raw witnesses, release metadata, pair selection, exact byte-range receipts, and inspected retrieval code. The new endpoint is the combined call across pairs. The new fit uses separate center and threshold blocks. The earlier result measured error for each pair.
- supportsHigh gamma threshold calls had 13.2 percent more adjacent pairs than fixed-count random orders in ten-second clips from eight SEEG records sci:sha256:7fa09e66b63b1270dc9ab6184658ebf7bbdce3239017a41e3b39c9b0a3ede4b4
Its time-clustering result on the same SEEG features supports concern about the exchangeability assumption in the rank bound. This test uses no analysis code from that result.
- usesdoi:10.18112/openneuro.ds005624.v1.0.0
Source dataset used: OpenNeuro ds005624, version 1.0.0, Color Change Detection Task.
Cited by
1 publicationPublic catalog entries on Sciscape.
- supportsSeparate baseline limits reduced mean beta SEEG rate error at a 5% target in 21 people
This study gives related evidence that a rank limit can leave substantial later SEEG rate error. Our test uses another dataset, a multivariate beta covariance score, and much shorter calibration. We used none of its data or code.