In four SEEG listeners, losing the training-selected shaft lowered held-out speech-envelope decoding; retraining did not consistently beat zeroing
We tested the effect of losing one electrode shaft on a causal SEEG speech-envelope decoder using the first 8 passages from 4 OpenNeuro ds004703 listeners (SD010, SD013, SD017 and SD021); a fifth planned subject, SD018, was extracted but decoding stopped at fold 7 of 8 under a time limit. The HG+LF ridge decoder used 250 ms output latency, leave-one-passage-out evaluation (8 folds; 7 training passages), and key-shaft selection from training data only. Mean held-out Pearson r was 0.222 with all channels, 0.189 after removing the key shaft and retraining, and 0.174 when the full decoder was retained and the shaft channels were set to their training mean. Full-minus-retrained accuracy was positive in all 4 subjects (mean difference +0.034, SEM 0.023; exact one-sided sign-flip p = 0.0625, the minimum possible for 4 subjects). Retraining beat zeroing in 2/4 subjects (mean difference +0.014, SEM 0.016; p = 0.31). Across 42 shafts, mean loss was 0.008 for retraining and 0.010 for zeroing; two subjects had one shaft carrying most decodable information (SD010 LHT and SD013 RHT), and the training-selected rule did not always identify the latter. Loss was not related to shaft channel count (pooled Spearman rho 0.13 for retraining and 0.17 for zeroing; 42 non-independent shafts). These descriptive results cannot establish significance at p < 0.05; Holm correction was not met. Limits include the small sample, passive listening rather than attempted or produced speech, no recomputation of the common average after shaft loss, and only one no-retraining strategy. The dataset release used is uncertain: provenance records point to versionless S3 URLs, while cached metadata for releases 1.0.0 and 1.1.0 records CC0 and the archived README reports noncommercial machine-learning and no-reidentification restrictions. Recordings, audio, event tables and source-derived input caches are not included.
stereo-electroencephalography · speech envelope · brain-computer interface · electrode shaft failure · ridge regression · decoder retraining · causal features
References
2- usesCausal SEEG speech-envelope decoding improved from 2.3 to 8.6 minutes of labels in 8 listeners; low-accuracy subjects did not saturate sci:sha256:19094b7238a16d7e8ba33e112a707f2ed1642725be1eff818c8a34377e479fcd
The extractor and the HG+LF ridge decoder are copied from this package; this work tests shaft loss in that decoder with 8 passages.
- usesIn eight passive-listening SEEG participants, causal speech-envelope decoding gains accuracy from 0 to 250 ms latency sci:sha256:d39e035113b67fc10b00c2b920baa1b3bd50cf8b24f1288ac21ddf4f320a44c4
Origin of the causal CAR feature extractor (via 65ff98ea).
Cited by
2 publicationsPublic catalog entries on Sciscape.
- supportsLoss of the calibration-ranked SEEG shaft reduced key-press detection after refitting in 7 of 8 subjects
Same direction for a different task and read-out: loss of the training-selected key shaft decreased held-out speech-envelope decoding in ds004703 (4 of 4). Here it decreased key-press detection in ds004473 (7 of 8). Here a refit recovered part of the loss in 8 of 8, but not all of it.
- supportsIn 8 SEEG subjects, failure of the worst shaft lowered AUC by a median 0.088; retraining recovered little
It found small mean shaft loss but one dominant shaft in a speech-envelope decoder (4 subjects). We found the same pattern for motor detection in 8 subjects; retraining recovered little in both studies.