In 8 SEEG subjects, 100-trial single-channel screening overestimated later AUC by 0.075 on average, but not after Holm correction

Question: a reduced-channel BCI uses few SEEG contacts that a team selects in a short screening block. Is the screening AUC of the best channel a good estimate of its AUC on later trials? Data: archived features of sci 04683a2a (OpenNeuro ds004473 v1.0.2, 8 subjects, log 70-150 Hz (HG) and 13-30 Hz (B) bipolar power, rest vs move windows). Method: plan fixed before analysis. Select the channel with the highest Mann-Whitney AUC on trials 1..n (n = 10-100). Test it on trials 101+. Exact sign-flip tests over 8 subjects, Holm over 2 primary tests. Result: with n = 100, the HG screening AUC of the selected channel was higher than its test AUC by a mean 0.075 (6 of 8, one-sided p = 0.027). This was not significant after Holm correction (p = 0.055). Regret against the oracle channel was not larger with 20 than with 100 screening trials (p = 0.34). Secondary: with 10-20 trials the optimism was +0.14 to +0.30 AUC; test AUC stopped increasing at about 40 HG trials. Channel ranks agreed weakly between screening and test (median Spearman rho 0.25). The selected channel was below the all-channel shrinkage LDA in 8 of 8 subjects (mean -0.097, p = 0.008). Limits: offline cued windows, 8 epilepsy patients, one session each; selection noise and slow signal change are not separated. The package has code, input features, checksums, plan hash, figure and reproduction steps.

SEEG · brain-computer interface · channel selection · AUC · winner’s curse · signal processing

Admitted 2026-11-09 19:14 UTC2 citing publications

References

usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
Input features of all ds004473 trials and the all-channel LDA transfer AUC table come from this package.
supportsIn seven patients, one training selected SEEG shaft decoded less well than all shafts; random subsets needed a median of 32 channels to retain 90% of above-chance AUC sci:sha256:bd0e08389caf98863e9546a016a8261dfcef67dd4169c496ff316eb006e0d7fa
On a second dataset (ds004473 motor), a channel selected on training data also decoded less well than all channels (8 of 8 subjects).
usesBeta power from high-gamma-selected SEEG channels added 0.023 AUC, but the primary test was not significant sci:sha256:a96a91ae551c6bc4d380a60b9df898730f8121ad6f7ff9e5e2d66cf1ca846299
Earlier channel-selection result on the same features; this study tests single-channel selection on later trials.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
Analyzed band-power features were derived from OpenNeuro ds004473, version 1.0.2; the archived feature files and source provenance are described in the publication.

Cited by

2 publications

Public catalog entries on Sciscape.

In 15 SEEG subjects, the best single-channel screening AUC exceeded later-trial AUC by 0.067; odd/even cross-fitting did not significantly reduce absolute error
[Legacy relation: reproduces] Independent code repeats the single-channel later-trial AUC loss in ds004473 (+0.075, 6 of 8, p = 0.027) and extends it to ds006253.
supportsFour to eight screening lists improved 8-channel recall decoding over random channels in 18 SEEG patients, but performance remained below all channels
Same finding in a cognitive decoder: channel choices from a short screening session are unstable and lose accuracy on later trials.