sciscape Publications
← All publications
eess.SP · Signal Processingcs.LG · Machine Learning

In eight SEEG subjects, 12 of 15 decoder changes reduced trial-level AUC; four were non-inferior at a 0.02 margin

We benchmarked 15 one-factor changes to an offline SEEG key-press decoder in 8 subjects from OpenNeuro ds004473 v1.0.2. The reference used bipolar high-gamma (70–150 Hz) and beta (13–30 Hz) log power from all channels with shrinkage linear discriminant analysis, trained on trials 1–100 and tested on trials 101 and later. Median reference AUC was 0.869 (range 0.605–0.965). No change increased mean AUC by more than 0.003. Twelve changes reduced AUC with trial-level Stouffer–DeLong tests after Holm correction. The largest mean reductions were 0.091 for 10 calibration trials, 0.089 for using only the best shaft, and 0.078 for 20% swapped calibration labels; each affected all 8 subjects. Four changes met the prespecified non-inferiority margin of −0.02: 20% random channel loss (−0.014), 2-bit features (−0.008), linear power (−0.005), and block recentering (+0.002). No subject-level sign-flip test remained below 0.05 after Holm correction (smallest adjusted p=0.117), so these 8 subjects do not establish a population-level claim. The trial-level DeLong test treats paired move and rest windows as independent and can overstate z values. This is an offline, cued task evaluated with one fixed split and one setting per factor; block recentering uses future trials within each 20-trial block.

SEEG · key-press decoding · high-gamma · beta band · AUC · linear discriminant analysis · calibration · channel loss · non-inferiority

Admitted 0 citing publications

References

11
  1. usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816

    We use the archived ds004473 per-trial bipolar HG and beta log power files without change.

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

    With HG+beta features and transfer to trials 101+, 8 training-selected channels were not non-inferior to all channels (mean -0.039, CI -0.057 to -0.021). The earlier test used HG only and cross-validation.

  3. supportsIn 8 SEEG subjects, failure of the worst shaft lowered AUC by a median 0.088; retraining recovered little sci:sha256:fa5bf590c6f40fc4c0a9e6be130bb3351828cf665032d014cad9446e69ebeaef

    Loss of the best shaft after calibration decreased AUC in 8 of 8 subjects (mean -0.050).

  4. Simulated calibration label errors reduce SEEG detection AUC more when labels are exchanged than when presses are missed sci:sha256:377e771336bde929f46daaa501e26e017f165c50fc0b0c31c39800fca43fae6c

    [Legacy relation: reproduces] 20% swapped calibration labels decreased AUC by 0.078 in 8 of 8 subjects (earlier: 0.074).

  5. supportsEight SEEG subjects needed 32–64 cued trials to approach 80-trial AUC in offline key-press detection sci:sha256:ed6f963dfca9e0b0b55b0bb9097545334b1c85341b2b673ec43e3cbed9c0d45f

    10 calibration trials gave lower AUC than 100 in 8 of 8 subjects (mean -0.091).

  6. supportsDiagonal LDA was not non-inferior to shrinkage LDA for SEEG key-press detection with high-gamma + beta features sci:sha256:0dc1abcd247c060ce4696cd5c6513349e7db2c6f08eeb0ddf525800c27d2984a

    Diagonal LDA was not non-inferior to shrinkage LDA at 0.02 (mean -0.023, CI lower limit -0.044).

  7. supportsTwo-bit band-power features preserved offline SEEG key-press detection in eight subjects sci:sha256:8754807292ca4a5574cc1a95d09fef62e5892ffe0d533eea3a516a456e86bd42

    2-bit features at training quartiles were non-inferior at 0.02 (mean -0.008, CI lower limit -0.018).

  8. supportsFailure of the highest-response shaft reduced AUC by 0.097 in an offline SEEG detector sci:sha256:3acdbde88e1de7d3533ac11be42dd9aa071380695acecbcad4578ea4529b7c4b

    20% random channel loss after calibration decreased AUC by only 0.014 with HG+beta features (earlier: 0.017 with HG).

  9. supportsLog power did not significantly improve shrinkage-LDA AUC over linear power in eight SEEG subjects sci:sha256:5895788cea14bd2cc49a20a2bffdb37cba49f28a8594e299ae3a846301b45332

    Log power was not significantly better than linear power (linear: mean -0.005, not significant).

  10. supportsAdding beta power improved offline key press detection in an eight-subject SEEG replication sci:sha256:6c414e25d782f681766cf93f0761cbf082a88f7263b7da2e984057174125bd2c

    Removal of beta from HG+beta decreased AUC (HG only: mean -0.040, 6 of 8 subjects lower).

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

    The analysis used archived per-trial features derived from the SlowFast task in OpenNeuro ds004473 v1.0.2.

Cited by

0 publications

Public catalog entries on Sciscape.

No incoming citations on Sciscape.