In 18 offline SEEG patients, refitting partly recovered AUC after 50% random channel loss; shaft loss caused greater decline
We tested a within-patient SEEG word-recall decoder under simulated channel loss after calibration, using archived features from OpenNeuro ds004789 v3.1.0, session 0, in 18 patients. At the primary 50% loss fraction, median held-out AUC was 0.681 with all channels, 0.660 after random loss with no refit, and 0.670 after refitting on stored calibration data. Exact two-sided Wilcoxon tests with Holm correction supported all three primary hypotheses: random loss with imputation reduced AUC (median paired difference -0.014, p = 0.0047); refitting improved AUC over imputation (+0.006, 16/18 patients, p = 0.0013); and refitting after shaft loss gave lower AUC than after matched-count random loss (-0.008, p = 0.0047). At 75% loss, refit median AUC was 0.588 after shaft loss and 0.642 after random loss. With a calibration-fixed threshold, imputation increased the median absolute flag-rate error from the 0.5 target to 0.083 at 50% loss and 0.142 at 75%; refitting kept it at 0.035 and 0.023. These results suggest that, for this offline decoder, retaining calibration data and refitting after channel loss can recover part of the AUC and maintain the flag rate, while shaft loss is more harmful than random loss. The study covers one task and session in one dataset; loss was simulated, stimulation was not tested, and z-scoring used all session words without labels. Four draws were run at each of two loss fractions because of the compute budget, rather than the ten draws at three fractions specified in the plan.
SEEG · word recall · memory BCI · channel loss · decoder robustness · logistic regression · AUC · calibration · intracranial EEG · neural decoding
References
2- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Uses the archived per-word SEEG band-power features, the inclusion rule and the within-patient decoder (C = 7.2e-4) of this publication.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
The archived features analyzed here are documented as coming from OpenNeuro ds004789 v3.1.0, FR1 session 0; the raw dataset was not downloaded again for this analysis.
Cited by
7 publicationsPublic catalog entries on Sciscape.
- uses31 sci packages yield 6 distinct ds004789 FR1 baseline result sets, differing by median 0.044–0.066 AUC per patient across 17 patients
We used column auc_full of results/per_patient.csv as the baseline AUC (family F2).
- supportsFailure of the highest-response shaft reduced AUC by 0.097 in an offline SEEG detector
Its finding that loss of whole shafts was worse than random loss also occurred in a motor detector: shaft failure minus same-size random failure was +0.002 on average (Holm p = 0.016) and +0.089 for the shaft with the largest training t.
- usesIn 18 SEEG patients, fixed-penalty logistic regression outperformed tuned and nonlinear recall decoders
Uses its all-channel median AUC (0.681, same features, model and folds) as a check of the reference decoder in this analysis.
- supportsSpreading eight SEEG pairs across shafts modestly improved word-recall decoding, but not within list position
With a different design on the same features (pair selection at a fixed count, not loss after calibration), we also found shaft-level redundancy: concentrated same-shaft pairs decoded worse than pairs spread over shafts, and same-shaft features were more correlated.
- supportsSimulated noisy SEEG channels slightly reduced recall-decoder AUC; repair did not improve the primary AUC test
It found that channel loss after calibration has a small AUC cost for this decoder. We found that noisy (not removed) channels after calibration also have a small AUC cost (0.001-0.005).
- supportsRecall decoding from SEEG increased with each added shaft up to the full implant in 18 patients
That work found that loss of 50% of channels after calibration decreased recall AUC by 0.014, and that loss of whole shafts was worse than random loss. We retrained the decoder on each subset and still found a loss when half of the whole shafts were removed (median 0.019, 18 of 18 patients). We also found that 2 whole shafts decoded worse than the same number of pairs spread over the implant, which agrees with the shaft-loss result.
- supportsDetecting simulated SEEG channel failures limited AUC loss more than refitting in a 15-patient working-memory decoder
Refit on surviving channels after channel loss recovered AUC in word recall (ds004789); this work shows the same in working-memory load decoding.