Fixed L2 penalty outperforms per-patient tuning for SEEG recall decoding in 18 FR1 patients
We tested whether per-patient tuning of the L2 penalty C improves within-patient recall decoders for SEEG. We used archived encoding-power features from 18 FR1 patients (OpenNeuro ds004789, session 0; source release not recorded), with log power at eight frequencies on each bipolar channel from 0 to 1.6 s after word onset. We compared balanced L2 logistic regression at the fixed RAM closed-loop value C = 7.2e-4 with C selected by inner cross-validation on training lists only from a seven-value grid (1e-5 to 1). In five outer list-block folds, tuned C gave lower AUC in 16 of 18 patients (median 0.635 vs 0.681; median difference -0.032; exact two-sided Wilcoxon p = 0.0028). AUC within serial position also decreased (median difference -0.020, p = 0.021). In a leave-one-list-out sensitivity analysis added after the primary result, tuned C was lower in 14 of 18 patients, with one tie (median difference -0.018, p = 0.011). The median AUC-versus-C curve was flat from 1e-5 to 3e-3 and peaked at 7.2e-4. Inner cross-validation selected all grid values, consistent with noisy selection. These findings suggest that, for this feature set, model, and training size, per-patient inner-CV tuning reduced held-out AUC; other tuning methods and effects on stimulation outcomes were not tested.
SEEG · recall decoding · logistic regression · regularization · cross-validation · AUC · brain-computer interfaces · memory
References
2- usesOne-list word-recall SEEG decoding mostly reflects serial position; three or more lists improve within-position AUC sci:sha256:6666745a5d262454b840ab848c437a2ab1e6d380dd1f88f42206228636891a04
That work lists the untuned penalty C as a limitation; this work tests per-patient tuning of C on the same type of decoder.
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Uses its 19 per-patient FR1 SEEG feature files (MD5 checked) and its model code; reproduces its W_chan leave-one-list-out median AUC 0.669.
Cited by
10 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_default of results/per_patient.csv as the baseline AUC (family F2).
- usesA session-0 word-recall decoder predicted session-1 recall above chance in 7 SEEG patients
We use the fixed penalty C = 7.2e-4, which this work showed to be better than per-patient tuning in FR1 SEEG.
- usesIn a four-patient RAM PAL1 pilot, study-phase SEEG band power did not predict cued recall consistently
We used the fixed RAM penalty C = 7.2e-4 because this work showed it is better than per-patient tuning in FR1.
- usesIn 18 FR1 SEEG patients, 10% simulated random label flips lowered median held-out AUC by 0.020; 20% exceeded the effective-sample rule
Fixed RAM penalty C = 7.2e-4, which this publication found better than a tuned penalty on the same data.
- usesLow-frequency SEEG features retained recall-decoding performance in 18 patients; high-gamma-only features performed worse
It supports the fixed penalty C = 7.2e-4 that we use for all frequency sets.
- usesIn one pyFR SEEG patient, the fixed-C spectral decoder did not exceed the permutation null (AUC 0.544, p = 0.15; n = 1; 3–140 Hz)
The fixed decoder penalty C = 7.2e-4 and the FR1 median AUC 0.681 reference line come from this publication. With n = 1, this pilot does not support or challenge it.
- usesIn 18 SEEG patients, an equal-weight cross-patient ensemble did not improve few-list word-recall decoding
Uses the fixed all-channel L2 penalty C = 7.2e-4 for W_k.
- usesA cross-patient prior did not improve recall decoding after 1–3 calibration lists in 18 SEEG patients
Uses the fixed penalty C = 7.2e-4 for the within-patient and prior-centred decoders, as that work found it better than per-patient tuning.
- usesIn 19 FR1 SEEG patients, a running quantile reduced offline stimulation-rate error
Fixed RAM penalty C = 7.2e-4 chosen because this work found it better than nested tuning.
- usesMatched-count left and right SEEG channels showed no detectable difference in word-recall decoding
Support for the fixed RAM regularization value C = 7.2e-4.