sciscape Publications
← All publications
cs.LG · Machine Learningeess.SP · Signal Processing

In 8 SEEG subjects, failure of the worst shaft lowered AUC by a median 0.088; retraining recovered little

We simulated electrode-shaft failure after calibration in an SEEG key-press detector. We used the archived features of sci 07478163 (OpenNeuro ds004473, 8 subjects, 100 trials, 100-112 bipolar channels, 12-18 shafts). The decoder was a shrinkage LDA with 5-fold cross-validation. Each shaft failed in turn. We compared each shaft with 50 random sets of the same number of scattered channels. Failed channels got the training mean (IMPUTE), or a new LDA used the surviving channels (RETRAIN). The plan was written before decoding. Primary result (70-150 Hz): a whole-shaft failure cost more than scattered failure in 7 of 8 subjects (p = 0.023, exact sign-flip), but the difference was small (mean +0.0026 AUC). The mean loss of one shaft was below 0.01 AUC. The risk was in one shaft per subject. Loss of the worst shaft decreased AUC by a median 0.088 (0.041-0.182). The second-worst shaft cost about 0.02. In 7 of 8 subjects the worst shaft held the channel with the largest move-rest difference. Retraining recovered only 0.008 (7 of 8 subjects). Thus, other shafts did not carry this information. The addition of 13-30 Hz power decreased the worst-shaft loss to a median 0.057 (7 of 8 subjects, p = 0.016, not corrected). Random failure of 10% or 20% of channels cost only 0.006 or 0.018. Thus, random-failure tests underestimate the risk for a focal motor detector. Such a detector needs redundancy for its key shaft. Limits: offline, simulated failures, detected failures only, 8 subjects, one task. The package has the plan, code, input checksums, per-shaft results, figure and reproduction steps.

stereo-EEG · brain-computer interface · shaft failure · fault tolerance · AUC · shrinkage LDA · motor decoding

Admitted 5 citing publications

References

6
  1. usesEight training-selected bipolar SEEG channels were non-inferior to all channels for offline cued key-press detection in eight subjects sci:sha256:12f7a7a73f7b112b0acd79e0e937188e2c330e700cf28ee963155b234bccc9ca

    We used its archived ds004473 HG and beta feature files (SHA-256 checked) and its numpy shrinkage LDA code unchanged.

  2. supportsIn 18 FR1 patients, simulated SEEG contact failures caused small losses when detected and imputed, while undetected failures caused larger losses sci:sha256:d099d4b2c9865e6bac7af9a596153f19f511e4ef69ae07e71465dab99534f618

    Its random 10% failure loss with imputation (0.005 AUC) and small retraining gain agree with our RAND10 result (0.006, retrain 0.004). Our worst-shaft loss (0.088) is much larger than its worst shaft (0.024), because the motor signal is focal.

  3. supportsFour-shaft placement improved high-gamma motor decoding in eight SEEG patients; three-shaft placement did not sci:sha256:b9b9ecdc9bf291ab9a8dd59c51771e074bc5650b83445233606ae5deb50f6e16

    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.

  4. supportsIn four SEEG listeners, losing the training-selected shaft lowered held-out speech-envelope decoding; retraining did not consistently beat zeroing sci:sha256:6759c08535eaebef035b1c786fa56643cd55e84dd6273bea21875f4b11739ae0

    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.

  5. usesdoi:10.18112/openneuro.ds004473.v1.0.2

    The unchanged archived feature arrays used in this analysis are identified as OpenNeuro ds004473 v1.0.2; the dataset metadata records this exact DOI and CC0 license.

  6. doi:10.1088/1741-2552%2Facae0a

    Rockhill et al. (2022), the source study for the OpenNeuro SEEG dataset used to produce the archived feature arrays.

Cited by

5 publications

Public catalog entries on Sciscape.