Per-shaft SEEG decoder scores were non-inferior to a joint decoder within 0.02 AUC in an eight-subject offline test
An SEEG implant can compute one decoder score on each depth shaft and send only that score. We tested if the plain sum of per-shaft scores detects key presses as well as one decoder on all channels. This sum is equal to an LDA with one covariance block per shaft. An earlier exploratory analysis on OpenNeuro ds004752 examined memory load; this preregistered test used a different dataset and task. Data were archived features from OpenNeuro ds004473 v1.0.2 (8 subjects; 70–150 Hz high-gamma and 13–30 Hz beta log power from 100–112 bipolar pairs; 12–18 shafts per subject). We used Ledoit–Wolf shrinkage LDA, with calibration on trials 1–100 and testing on trials 101 and later. The plan was preregistered. Mean test AUC was 0.825 for the sum of shaft scores and 0.828 for the joint decoder. The sum was non-inferior within a 0.02 AUC margin (mean difference −0.004; 3 of 8 subjects higher; exact sign-flip p = 0.020). A second-level LDA on shaft scores was also non-inferior (mean difference +0.010; 6 of 8 higher; Holm p = 0.012). The best single shaft was lower by 0.039 (7 of 8 lower; Holm p = 0.078, not significant). After loss of one shaft without refit, no difference between the sum and joint decoder was shown (mean difference −0.001; Holm p = 0.84). The modular decoder sends 12–18 values per window instead of 200–224 features. In these data, cross-shaft covariance gave almost no decoding value. Limits: offline analysis of one dataset and task, 8 subjects, one linear decoder, same-session testing, and no quantization or hardware test. Subject 11 had only 32 test trials, and subject 12 was near chance (AUC about 0.6).
SEEG · intracranial EEG · brain-computer interface · linear discriminant analysis · shrinkage LDA · decoder fusion · non-inferiority · OpenNeuro · preregistered analysis
References
4- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
We use the archived ds004473 HG and beta features (hashes checked) and the unchanged lda_np.py of this package.
- supportsDiagonal LDA was not shown non-inferior for high-gamma and beta decoding and lost AUC with five bands in 15 SEEG patients sci:sha256:975c5ada4b10cda76eb29d70ffcfdcd7a6f9f07cd9ea79524f938750e4079064
Its exploratory result (one covariance block per shaft equal to full shrinkage LDA, ds004752) is supported by a preregistered test on a different dataset and task (ds004473).
- usesIn 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
We use its one-shaft loss design with mean fill as the joint-decoder comparator; our joint value agrees with its value.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses archived high-gamma and beta features derived from OpenNeuro ds004473 v1.0.2; the source package records raw-file URLs, sizes, and S3 ETags.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- Summed per-shaft SEEG scores were non-inferior to joint decoding, while four selected shafts were not shown non-inferior in 15 patients
[Legacy relation: reproduces] Same SUM vs JOINT non-inferiority test on a second dataset and task; result reproduced (-0.004 AUC on both).
- supportsSummed per-shaft SEEG scores were non-inferior for word-recall decoding, with a small AUC loss (ds004789 FR1, 18 patients)
Its motor-task result (sum of shaft scores non-inferior to joint decoder) also holds for word-recall decoding at the same 0.02 margin, with a similar small mean loss.