A Gaussian kernel did not improve SEEG word recall prediction over a linear decoder in 18 patients
We compared two within-patient decoders of later word recall. The data are SEEG encoding power from 18 patients in OpenNeuro ds004789 (FR1 free recall, session 0). The features are log10 power at 8 frequencies (3-140 Hz), averaged 0-1.6 s after word onset, on all bipolar channels. Both decoders are kernel L2 logistic regression with leave-one-list-out cross-validation and balanced class weights. The linear decoder (LIN) uses a linear kernel. The primary nonlinear decoder (RBF1) uses a Gaussian kernel with the median-heuristic width, and its regularization matches LIN in the near-linear limit. We wrote the plan before the fits. Result: median AUC 0.675 (RBF1) vs 0.669 (LIN). The median paired difference is +0.0019 (95% bootstrap CI -0.0018 to +0.0047; 10 of 18 patients higher; exact two-sided Wilcoxon p = 0.28). The CI is inside +-0.02, so the two decoders are equivalent within 0.02 AUC. A narrower kernel (-0.007), a wider kernel (+0.001), the AUC within serial position (+0.0005), and an RBF decoder on 8 channel-mean features (-0.006) also showed no significant difference. Channel-level detail (LIN vs channel-mean LIN) adds +0.055 AUC (p = 0.0016), which is much more than the kernel adds. Limits: time-averaged features only, about 300 words per patient, no nested tuning of the kernel width or C, and one dataset. Post-hoc, the RBF1 and LIN scores have a median correlation of 0.986, so RBF1 is close to linear in this regime. The package contains the input features with provenance, numpy-only code, the plan, per-patient results and a figure.
SEEG · word recall · memory decoding · intracranial EEG · Gaussian kernel · RBF · linear decoder · logistic regression · AUC · leave-one-list-out cross-validation
References
3- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Input features (ds004789 FR1, 18 patients), inclusion rule, linear W_chan decoder settings and statistics helpers come from this publication. Our LIN decoder reproduces its W_chan median AUC (0.669).
- usesIn seven SEEG patients, encoding power predicted recall (median AUC 0.686), while task-adjusted within-position AUC was not significant sci:sha256:751ab98c23e7a3dae98de88344498180d3d994c09700abdef430fccd5cee1579
Feature-extraction code (fetch_features.py) and 7 of the 18 patient feature files come from this publication, through c4b8c884.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
The analyzed input features are from OpenNeuro ds004789 v3.1.0, RAM FR1, session 0, acq-bipolar.
Cited by
5 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_LIN of results/per_patient.tsv as the baseline AUC (family F1).
- supportsDefault nonlinear decoders did not improve memory-load decoding over shrinkage LDA from SEEG band power
A Gaussian-kernel decoder was again not better than a linear decoder, here for memory load in ds004752.
- usesThe decoder comparison stopped after one patient and provides no group result for memory load from SEEG band power
Prior result for FR1 word recall: a Gaussian-kernel decoder was not better than a linear decoder.
- supportsDefault nonlinear decoders did not outperform shrinkage LDA for SEEG working-memory load in 15 patients
A Gaussian-kernel decoder gave no gain over a linear decoder for word recall; RBF-SVM also gave no gain for memory load here.
- supportsIn 15 SEEG patients, nonlinear decoders did not outperform shrinkage LDA for offline memory-load decoding
Gives the same result (no nonlinear gain over linear decoders) in a different SEEG task (Sternberg memory load, ds004752).