In 8 SEEG subjects, label-free imputation matched a labeled refit after known channel loss, but key-shaft loss still reduced AUC
We tested whether conditional-mean imputation (CMI), which uses no labels, can replace a labeled refit after simulated loss of known SEEG channels. Archived features from OpenNeuro ds004473 v1.0.2 (CC0; Rockhill et al. 2022) covered 8 subjects performing a cued key press versus rest. A fixed shrinkage LDA used log 70–150 Hz high-gamma and 13–30 Hz beta power, trained on trials 1–100 and tested on trials 101 onward. A plan was written before decoding. After loss of the key shaft, CMI improved AUC over mean fill by 0.017 (7 of 8 subjects; exact sign-flip p = 0.0078, Holm p = 0.012). CMI was non-inferior to a labeled refit within a 0.01 AUC margin after key-shaft loss (mean CMI − REFIT = −0.0003; 8 of 8; Holm p = 0.012) and random loss of 25% of pairs (−0.0007; 8 of 8; Holm p = 0.012). This confirms an exploratory result with new code and a planned test. Yet CMI and REFIT recovered only about 25% of key-shaft AUC loss: mean loss was 0.068 with mean fill and 0.052 with CMI; the maximum remaining loss was 0.135. A refit after known channel loss may not need new labels for this linear decoder, but software did not replace the key-shaft information. This offline study simulated loss in one session, used one task and decoder, and included 8 subjects; the key shaft was selected by high-gamma t-values only.
SEEG · brain-computer interface · conditional-mean imputation · channel loss · shrinkage LDA · AUC · high-gamma · beta power
References
4- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
Input features (ds004473, all trials, 8 subjects) and the shrinkage-LDA code lda_np.py; checksums agree.
- supportsIn 8 SEEG subjects, failure of the worst shaft lowered AUC by a median 0.088; retraining recovered little sci:sha256:fa5bf590c6f40fc4c0a9e6be130bb3351828cf665032d014cad9446e69ebeaef
Key-shaft loss is the main risk; with HG+beta, a labeled refit or label-free CMI recovered only 0.017 of a 0.068 AUC loss.
- supportsIn 8 offline SEEG subjects, dropout augmentation lowered AUC; conditional-mean imputation was within about 0.001 AUC of refits at 25% and 50% simulated channel loss sci:sha256:3a2412f50543607ab1a2d73aeba29dce9596abcd5a9d9de95cc4b2bdce60f9ac
Planned confirmatory test of its exploratory result: CMI is non-inferior to a labeled refit within 0.01 AUC at 25 % random loss (8 of 8, Holm p = 0.012), with new code.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed archived feature files are identified as features from OpenNeuro ds004473 v1.0.2; their checksums match the source contribution.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- Label-free imputation matched refitting after known SEEG channel loss, but recovered only 16–22% of key-shaft loss in 15 patients
[Legacy relation: reproduces] Same pre-registered design (key shaft and random 25 % loss; fill, CMI, refit; margin 0.01) in a second dataset and task. H1-H3 pass again; CMI recovered 16-22 % of the key-shaft loss here vs about 25 % there.
- Loss of the most informative shaft reduced movement decoding AUC by 0.061 in 15 ds006253 SEEG patients
[Legacy relation: reproduces] Same key-shaft definition and CMI/mean-fill/refit comparison. ds006253 gives a 0.061 mean-fill and 0.051 CMI cost, against 0.068 and 0.052 in ds004473; CMI is again within 0.01 of a refit.
- Conditional-mean imputation matched refitting after simulated SEEG channel loss but recovered only 14% of key-shaft AUC loss
[Legacy relation: reproduces] Same 3 pre-registered hypotheses (H1 KEY CMI > MEAN; H2 KEY CMI non-inferior to REFIT; H3 R25 CMI non-inferior to REFIT) and the same lda_fit code, on a second dataset (ds006253, 15 patients, mouse movement) with HG only. All 3 pass. The fraction of the key-shaft loss that CMI recovers is lower here (14% vs about 25%).