sciscape Publications
← All publications
q-bio.NC · Neurons and Cognition

Decoder selection explained most of the lower recall of stimulated words in RAM closed-loop memory experiments

Question: how well did the online decoder of a deployed closed-loop memory BCI separate good and poor encoding? In RAM FR3 and catFR3, a logistic classifier triggered stimulation when its output was below the training median. Stimulated words had lower recall than unstimulated words in the same lists (0.205 vs 0.332; D_obs = +0.127, 95% CI 0.083 to 0.176, 19 of 20 subjects, sign-flip p = 1e-5). This contrast mixes decoder selection and the stimulation effect. We used the randomized NoStim lists to separate them. Assumptions: stimulation does not change recall of unstimulated words in the same list, and list type is random. Then the decoder separation without stimulation is D0 = (pU - pN)/f. D0 was +0.090 (CI -0.014 to 0.190; 16 of 20 positive, sign test p = 0.012, sign-flip p = 0.11). A ratio-of-means estimator gave +0.114 (CI 0.033 to 0.198). The implied stimulation effect on stimulated words was -0.036 (CI -0.116 to 0.026), not significant, near the open-loop estimate of sci 4d5c3ad0. A binormal decoder with AUC 0.65 would give D0 = 0.158. The implied online AUC is 0.59 to 0.61 (CI 0.50 to 0.69), lower than offline FR1 AUC 0.67, but the CI includes 0.65. After serial-position adjustment, D0 fell to +0.050 (not significant): stimulation rate was 0.33-0.39 at positions 1-2 and about 0.6 at positions 6-12. The stimulation rate per session was 0.23 to 0.80 and differed from 0.5 in 15 of 35 sessions. No spillover to the next word was found (-0.005, p = 0.90). Results held with 4 early-latency sessions added (22 subjects) and in FR3 alone. Limits: no classifier outputs in the files, so AUC is model-based; D0 is noisy when f is small; ECoG and SEEG are mixed; 20 subjects. The package has code, file manifest with SHA-256, tables, figures and reproduction steps.

closed-loop brain-computer interface · memory · intracranial stimulation · decoder · recall · serial position

Admitted 3 citing publications

References

4
  1. supportsOpen-loop stimulation during word encoding reduced recall by 1.6 points overall and 4.3 points at hippocampal sites sci:sha256:3a50433ef645749679239857437f277c2d17f9ae681ef4ff329341dc5afe1650

    Independent per-word analysis of the same closed-loop sessions gives a list-level stimulation effect near zero (-1.2 points), consistent with its closed-loop result.

  2. usesOpen-loop stimulation during word encoding reduced recall by 1.6 points overall and 4.3 points at hippocampal sites sci:sha256:3a50433ef645749679239857437f277c2d17f9ae681ef4ff329341dc5afe1650

    We use its finding that unstimulated words in stimulation lists did not differ from no-stimulation lists (-0.2 points) to support assumption A1. Our implied stimulation effect on stimulated words (-2.2 to -3.6 points, not significant) and closed-loop list effect (pL - pN = -1.2 points) agree with its open-loop (-1.6) and closed-loop (-1.4) estimates.

  3. supportsOne-list word-recall SEEG decoding mostly reflects serial position; three or more lists improve within-position AUC sci:sha256:6666745a5d262454b840ab848c437a2ab1e6d380dd1f88f42206228636891a04

    In the deployed RAM closed-loop classifier, the stimulation rate depends strongly on serial position and the position-adjusted decoder separation is smaller and not significant. This agrees with its finding that much of a small-data recall decoder's signal comes from serial position.

  4. usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a

    We compare the online decoder separation with the offline within-patient FR1 decoder AUC (median 0.669) from this publication, through a binormal reference curve.

Cited by

3 publications

Public catalog entries on Sciscape.