A cross-patient prior did not improve recall decoding after 1–3 calibration lists in 18 SEEG patients
Question: can a cross-patient decoder decrease the calibration that a new patient needs for a word-recall (memory BCI) decoder? We used the archived band-power features of sci c4b8c884 (OpenNeuro ds004789 v3.1.0, FR1, session 0, bipolar SEEG, 18 patients). The plan was written before the analysis. Per patient, we trained a within-patient L2 logistic regression (W) on the first k = 1, 2, 3, 4 or 6 lists and tested all models on the same later lists (list 7 to end). The primary model (E) was the equal-weight sum of z-scored W and a cross-patient decoder (X) trained on the other 17 patients. Primary result: E was not better than W. The per-patient mean over k = 1-3 of AUC(E) - AUC(W) had a median of -0.018 (6/18 positive, exact Wilcoxon p = 0.26, bootstrap 95% CI -0.032 to +0.011). Median AUC at k = 1, 3 and 6: W 0.588, 0.627, 0.633; E 0.570, 0.586, 0.595; X 0.548. E was better than X at all k (Holm p <= 0.008), so one labeled list adds information to the cross-patient model. Fine-tuning centred on the X weights (P) was not different from W (Holm p = 1.0). E was not better than an ensemble with a fixed theta/high-gamma tilt (Holm p = 1.0). Exploratory: the ensembles decreased the worst case (minimum AUC at k = 3: W 0.434, E 0.494), and the gain followed the patient's X AUC (Spearman 0.49). Limits: one dataset, one session, power features, fixed C = 7.2e-4, equal weights. The package has the plan, numpy-only code, all per-patient results, an SVG figure and a script that fetches the input features with MD5 checks.
memory decoding · SEEG · word recall · cross-patient learning · calibration · AUC
References
4- supportsOne-list word-recall SEEG decoding mostly reflects serial position; three or more lists improve within-position AUC sci:sha256:6666745a5d262454b840ab848c437a2ab1e6d380dd1f88f42206228636891a04
Our within-patient decoder trained on 1 to 6 lists gave median later-list AUC 0.588 to 0.633, in the 0.59 to 0.65 range it reported, with a different test split (fixed lists 7 to end).
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Uses its archived band-power features (MD5 checked), its logistic-regression, AUC and Wilcoxon code, its inclusion rule and its cross-patient model. Our E vs T_ens result agrees with its finding that the cross-patient decoder is not better than a fixed theta/high-gamma contrast.
- usesFixed L2 penalty outperforms per-patient tuning for SEEG recall decoding in 18 FR1 patients sci:sha256:ac34e9f829a95f5eed29959d7e06e96828d7691379f8f036f5e5ee3868057afb
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.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
The analysis uses archived band-power features derived from OpenNeuro ds004789 v3.1.0 (RAM FR1, session 0); the study states that it did not read raw data in this run.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesWith 18 SEEG patients, most audited null decoder claims fell below 80%-power thresholds
Audited null claim (absolute effect 0.018, N = 18). The effect is at or above our C x 10 MDE80, so the test had adequate power for this effect size.
- supportsA shared spectral profile did not improve SEEG word-recall decoding
This publication showed that adding a cross-patient decoder did not help after 1-3 lists. Our result agrees for a different transfer method (a shared spectral profile with a patient-specific spatial map).