Within-patient SEEG recall-decoder AUC continued to rise through 240 words; a group-shift forecast matched a power law in 18 patients
We measured within-patient word-recall decoder learning curves in 18 patients with stereo-electroencephalography (SEEG), using archived FR1 band-power features from session 0 of OpenNeuro ds004789 v3.1.0. We used five-fold list-blocked cross-validation and training pools of 1–20 lists (12–240 words), with fixed L2 logistic penalties. The all-channel decoder (W_chan; C = 7.2e-4) had median AUCs of 0.527 after 12 words, 0.595 after 48 words, and 0.670 at the full pool. From 120 words to the full pool, AUC increased in 16 of 17 patients with more than 10 lists (p = 1.5e-5); the median log-linear gain was about 0.031 AUC per doubling. The fitted power-law exponent was small (median 0.11), and 9 of 18 fitted asymptotes exceeded 1, so one session did not identify the asymptote and no saturation word count is reported. To predict full-pool AUC from the first four lists, a power law with a group exponent had median absolute error 0.029, better than no gain (0.072, p = 0.0013), but not better than adding the other patients’ median gain (0.023, p = 0.44). An eight-feature channel-mean decoder had similar AUC after 12 words and lower AUC at the full pool (0.595 vs 0.670, p = 8.4e-4). Limits include one task and session, fixed penalties, random rather than time-ordered training lists, and optimistic forecast errors because short-calibration and full-pool AUCs shared test words. This is an offline analysis; it did not test stimulation or closed-loop decoding.
stereo-electroencephalography · SEEG · free recall · memory decoding · learning curves · area under the curve · logistic regression · calibration · cross-validation
References
4- usesIn 18 SEEG patients, nonlinear decoders did not outperform a linear decoder for word recall from band power sci:sha256:25ddfb21f49286c69b9c9c27135bc98bf4f572b3f2f04922e1b58c80c094f3fb
Uses its fixed linear penalty C = 7.2e-4; its finding that nested tuning decreased AUC is the reason that we kept C fixed for all training sizes.
- supportsCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Extends its calibration result (W_chan better than a cross-patient decoder after 3 lists): within-patient AUC continues to increase to 240 words and does not saturate in one session.
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Uses the archived FR1 SEEG feature files, the inclusion rule, the decoders and the logistic solver of this publication.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
The archived FR1 session-0 features analyzed here are identified in the publication as coming from OpenNeuro ds004789 v3.1.0.
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesEarly power-law learning curves did not improve 100-trial SEEG decoder forecasts over cross-validation in eight subjects
Prior learning-curve result (AUC gain per doubling of training data) that motivated the test of power-law extrapolation.