In 15 SEEG patients, simulated gain drift lowered load-decoding accuracy, while spectral-shape features preserved balanced accuracy

Question: a chronic BCI channel can change its gain, for example when the electrode impedance changes. Does this drift decrease an SEEG band-power decoder of working-memory load? Do features that remove the per-channel broadband level keep the accuracy? Method: we used the archived log10 band power of 15 patients (5 bands, 3-6 s maintenance, set size 8 vs 4, 20-42 trials) and shrinkage LDA with a fixed threshold in 5-fold cross-validation (10 repeats). We trained on clean data. Then we added a random gain per channel (SD 0, 1, 3 or 6 dB) to the test trials. Results: the full decoder decreased from balanced accuracy 0.649 to 0.567 at 3 dB (p = 0.003) and to 0.535 at 6 dB. Spectral-shape features (log power minus the channel mean over bands) are exactly invariant to the gain. They gave 0.649 at all drift levels. Without drift they were equal to the full features (difference +0.000, 95% CI -0.049 to +0.049). With 3 dB drift they were better (+0.083, 13/15 patients, exact sign-flip p = 0.0004). AUC changed less than balanced accuracy, so most of the loss is a threshold shift. Label-free re-centring of the full decoder on the test batch was also invariant (0.678). It was not significantly better than the shape features (p = 0.27). Limits: the drift is simulated and constant, the decoding is offline and within one session, and re-centring used class-balanced test batches.

SEEG · working memory · band power · gain drift · spectral shape · shrinkage LDA · balanced accuracy

Admitted 2026-11-09 18:40 UTC4 citing publications

References

usesAll-trial feature selection raised SEEG decoding AUC to 0.84 with random labels sci:sha256:d0f8e2f61f20ff7561d81bf30c1c96cfde2212b5dc53c421ac4d3aefa770c0ad
We copied the 15 archived band-power feature files (data/feat, SHA-256 verified), the trial rule (valid, artifact 0, set size 4 or 8), and the shrinkage LDA and AUC code from this package. Its all-feature baseline (AUC 0.70) agrees with our no-drift full decoder (AUC 0.706).
usesIn 15 SEEG patients (OpenNeuro ds004752), a band-power decoder trained only on memory loads 4 and 8 puts load 6 at the midpoint, but its output does not predict response time within a load sci:sha256:e2745563d558339564cbc0b329f8efc977263b7a2e3956a01d641ce57864f0b3
This package archived the band-power features that we used. 7ff3bfd4 copied them unchanged.
usesBetween-session SEEG band-power shifts exceeded within-session shifts in 7 of 8 patients, with unknown session dates sci:sha256:bf4ff716f0799dd68007810406a9a856f04bc3e6069321f8f3be767be8b0b17d
This result showed that SEEG band-power features shift between recording sessions (ds005624). That finding is the reason we tested a simple model of such a shift (a per-channel gain change) and features that are invariant to it.
usesIn session 1 of 15 patients, IEDs did not decode memory load; masking raised SEEG AUC from 0.70 to 0.72 sci:sha256:632c406c34e1530eea3bb23516da5054eb6c64dbaa63783d6d4628a34fe21bbb
This package computed the bipolar log10 band-power features from the raw OpenNeuro ds004752 SEEG. Our data/provenance.json comes from this chain.
doi:10.7554/eLife.78677

Cited by

4 publications

Public catalog entries on Sciscape.

supportsSpectral-shape features preserve SEEG word-recall decoding under simulated gain drift, with little cost without drift
Same decoder and drift code, applied to a second SEEG task (word recall, 18 patients). Shape features were invariant and non-inferior at 0 dB (AUC -0.003, CI -0.018 to 0.012). Shape was better than full at 3 dB (+0.018 balanced accuracy, p = 0.047) and at 6 dB (+0.031, p = 0.00085). All effects had the same direction, but they were approximately 4 times smaller than for working-memory load.
supportsIn ds004789 FR1 data, spectral-shape features retained word-recall decoding accuracy under simulated per-channel gain drift
The same result (gain drift harms band-power decoding; gain-invariant spectral-shape features keep accuracy) holds for a second task and dataset: FR1 word-recall decoding in ds004789.
supportsShape features preserved fixed-threshold SEEG key-press accuracy under 3 dB gain drift but reduced AUC without drift in eight subjects
Same direction for a motor detector: gain-invariant shape features keep fixed-threshold accuracy under simulated per-channel gain drift, as shown for working-memory load.
challengesGain-invariant spectral features did not improve cross-session SEEG accuracy, while session centring did
This work tests the simulated-gain-drift claim on real between-session drift: shape features gave no gain in cross-session balanced accuracy, but label-free centring did.