Four-shaft placement improved high-gamma motor decoding in eight SEEG patients; three-shaft placement did not
This study tested whether placing bipolar channels on different shafts improves decoding of a focal motor signal. We used OpenNeuro ds004473 v1.0.2 (SlowFast task), with eight patients and the first 100 trials per patient. For each patient and k = 2, 3, or 4, we compared 100 random sets of k consecutive bipolar rows on one shaft (ADJ) with 100 sets of one row on each of k different shafts (SPREAD). The pre-specified primary analysis used log 70–150 Hz power, shrinkage LDA, five-fold cross-validation, and exact two-sided sign-flip tests with Holm correction for k = 3 and 4. At k = 4, SPREAD exceeded ADJ by a median 0.021 AUC (7/8 patients; sign-flip p = 0.016; Holm p = 0.031). At k = 3, the median difference was +0.008 AUC and was not significant (p = 0.109; Holm p = 0.109). Secondary 13–30 Hz comparisons favored SPREAD at k = 3 and 4 (median differences +0.024; Holm p = 0.047). The signal was focal: 0–8% of rows per patient had strong single-row decoding, across 0–3 shafts. Random sets of 2–4 rows therefore had low mean AUCs (0.53–0.59), while the all-channel median AUC was 0.840. Median within-class noise |r| at 70–150 Hz was 0.31 for adjacent rows, 0.08 for non-adjacent rows on one shaft, and 0.06 across shafts. The best of 100 draws was similar for ADJ and SPREAD at k = 4 (0.793 and 0.786; descriptive only), although the number of possible sets sampled differed. The result supports shaft-spread placement for random placement when the motor focus is unknown; it does not establish the best placement for implants targeting a known focus. Limitations include eight patients, one task, one session, clinical electrode placement, offline decoding, no closed-loop BCI test, secondary status of the beta-band analysis, and no nested selection using training data.
stereo-EEG · motor decoding · electrode placement · high-gamma power · shrinkage LDA · OpenNeuro ds004473
References
6- supportsThree to four SEEG channels on different shafts improved recall decoding over channels on one shaft in 19 patients sci:sha256:558fe6f889fb23a1e9f9bc942ba8ba7afbf19a1b390edcdc000ef79fac9b9868
Extends the shaft-spread result from word recall (ds004789) to key press vs rest; same direction and similar size at k = 4.
- usesSparse SEEG contacts showed an inconclusive small read-out decrease, with no detected penalty at equal channel count sci:sha256:f758ddd6a1f33f92c859c7b0dd2e5465eeb094a9e5d7de64d8afe4381926adbb
Reuses its ds004473 trial windows, features, data reader, and LDA code without change; the FULL 70-150 Hz median AUC (0.840) reproduces its value.
- supportsSpreading eight SEEG pairs across shafts modestly improved word-recall decoding, but not within list position sci:sha256:82a78f4f7dc97abeff3d496e93778d71be2aeb945dc4fd650b94278827205191
Same mechanism: adjacent rows on one shaft have much higher within-class noise correlation than rows on different shafts (0.31 vs 0.06 at 70-150 Hz).
- supportsThree to four SEEG channels on different shafts decoded working-memory load better than adjacent channels on one shaft sci:sha256:fa3a53e1ec5646b27db6a0b0e1fbefe1ff41dbc0d252ce7e2a5d75bb6958cdfb
Same SPREAD vs ADJ design; the k = 4 gain also occurs for a focal motor signal in a different dataset (ds004473), while k = 3 was not significant at 70-150 Hz.
- usesRe-referencing improved high-frequency SEEG detection over the recorded reference, but no montage was best across eight subjects sci:sha256:fde123191b6f52b597384fb536814e8db9f51ec1854910522f07a26b58c04704
The EDF byte-range reader and numpy shrinkage LDA come from this package.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The study analyzed the SlowFast task in OpenNeuro ds004473 v1.0.2.
Cited by
5 publicationsPublic catalog entries on Sciscape.
- With training-fold selection, four-shaft SEEG sets showed no clear high-gamma gain over adjacent pairs
[Legacy relation: reproduces] Part A repeats the random-placement SPREAD vs ADJ test at k = 4, 70-150 Hz with new code, other windows and 60 trials. The direction and size agreed (+0.018 vs +0.021 AUC, 7 of 8 subjects), but Holm p was 0.063. Part B adds nested set selection, which removed the gain.
- challengesRandom SPREAD placement did not outperform DENSE placement for movement decoding in seven SEEG patients
The original SPREAD > DENSE finding at K = 4 (ds004473) did not generalize to ds006253 (+0.005 AUC, 3/7 patients).
- supportsFailure of the highest-response shaft reduced AUC by 0.097 in an offline SEEG detector
It found that channels on several shafts decoded movement better than channels on one shaft. Our failure analysis is consistent: the motor information is concentrated on one shaft, and its loss decreased AUC by 0.097, so spread over shafts gives redundancy.
- supportsSpreading SEEG bipolar pairs across shafts improved key press decoding at 4, 8, and 16 pair budgets in eight subjects
We find the same spread benefit at 4 pairs with HG+beta features and show that it continues and increases at 8 and 16 pairs.
- supportsIn 8 SEEG subjects, failure of the worst shaft lowered AUC by a median 0.088; retraining recovered little
It showed that key-press information is on few shafts. We found that one shaft per subject carries most of it: its loss cost 0.04-0.18 AUC, and the second-worst shaft about 0.02.