In ds004789 FR1 data, spectral-shape features retained word-recall decoding accuracy under simulated per-channel gain drift
We tested if a gain-invariant feature set keeps a SEEG memory decoder accurate when the gain of each channel changes. Data: archived log10 power features of sci c4b8c884 (OpenNeuro ds004789, FR1, session 0, bipolar, 8 frequencies 3-140 Hz, 0-1.6 s after word onset; 18 patients). Task: predict later recall of each word. Decoders: L2 logistic regression and shrinkage LDA, 5-fold cross-validation grouped by list. We trained on clean data and added one random gain per channel, N(0, sigma) dB, to the held-out words (sigma 0, 1, 3, 6 dB; 20 repeats). The plan was written before the analysis. Results (logistic regression, exact Wilcoxon over patients): (1) Drift decreased log-power (FULL) balanced accuracy from 0.609 to 0.573 (median) at 3 dB (mean diff -0.029, 95% CI [-0.043, -0.015], p = 0.0013) and to 0.546 at 6 dB. (2) Without drift, shape features (log power minus the channel mean over frequency) were non-inferior to FULL (AUC diff -0.002, CI [-0.008, 0.003], margin 0.02). (3) At 3 dB, shape features had higher balanced accuracy than FULL (+0.028, CI [0.016, 0.040], p = 0.00025, 15 of 18 patients). LDA gave the same three results. Exploratory: the broadband level alone has less information than shape (AUC -0.038, p = 3.8e-5). Label-free re-centring of FULL on the new-session words is also exactly invariant and equal to shape, but it needs a batch of unlabelled words. Limits: the drift is simulated, pure gain, constant over frequency and time; one session per patient. This extends the ds004752 result of sci 10e41ff5 to a second task and dataset. Sci be64689c found no real cross-day benefit of shape features on ds004752. Thus our simulated benefit is an upper bound for drift that is mostly per-channel gain. The package has code, input features with hashes, per-patient results, figure and instructions.
SEEG · word recall · neural decoding · spectral shape · gain drift · subsequent memory · brain-computer interface
References
4- supportsIn 15 SEEG patients, simulated gain drift lowered load-decoding accuracy, while spectral-shape features preserved balanced accuracy sci:sha256:b6263c39028f94a46261aa5c8e9fe09fbb227be48cf4fbdab0053e242c0234aa
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.
- supportsWithin-session SEEG decoding showed spectral shape carried working-memory load, but non-inferiority to full band power was not established sci:sha256:18e08468d6eb44fdb469c044e816472c235c81d6cb5820ccd2bcb931abfa4aea
Shape features without the broadband level keep most decoding information; here shown non-inferior to full log power at a 0.02 AUC margin for FR1 recall decoding (ds004789).
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Input features (ds004789 FR1 log10 power, 18 included patients) are copied from this publication; SHA-256 values match.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
The analyzed ds004789 FR1 features were copied from the cited internal feature contribution and match its recorded hashes; the source publication identifies OpenNeuro ds004789 v3.1.0.
Cited by
1 publicationPublic catalog entries on Sciscape.
- Spectral-shape features preserve SEEG word-recall decoding under simulated gain drift, with little cost without drift
[Legacy relation: reproduces] Concurrent independent analysis of the same question on the same 18 patients and features, with shrinkage LDA, 5 list-grouped folds, 5 repeats and sign-flip tests. It gives the same three conclusions: shape non-inferior at 0 dB, shape better than full at 3 dB, full decreases under drift. Difference: the full balanced-accuracy drop at 3 dB here was smaller (-0.019, p = 0.061, LDA) than in 595acd34 (-0.029, p = 0.0013, logistic regression); the AUC drop here was significant (-0.031, p = 0.0002).