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

An adaptive threshold reduces stimulation-rate error in an offline memory-decoder simulation

A closed-loop memory brain–computer interface (BCI) can stimulate when a recall decoder predicts poor encoding. We tested whether a threshold calibrated at the start of a session maintained a target flag rate of 0.5 later in that session, using archived FR1 stereoelectroencephalography features from 18 patients. An L2 logistic decoder (C = 7.2 × 10⁻⁴) was calibrated on lists 1–8 with a threshold set from leave-one-list-out outputs. On lists 9 onward, the fixed threshold flagged a median 0.589 of words (IQR 0.513–0.688); rates were above 0.5 in 14/18 patients (exact two-sided Wilcoxon p = 0.0046), with median absolute error 0.101. The shift was present from the first test list, without a significant rate slope (p = 1.0), and decoder outputs were a median 0.28 calibration standard deviations below the threshold. Across 50 random list-order shuffles, group median rates ranged from 0.46 to 0.55 and none reached 0.589. A label-free threshold set to the median decoder output over the previous four lists gave a median flag rate of 0.510 and median absolute error 0.028 (p = 0.0026 versus fixed). Decoder scores and AUC were unchanged (median test AUC 0.621; 0.580 within serial position). Re-z-scoring features from the previous four lists reduced median absolute rate error to 0.043 but lowered AUC (p = 0.005). With 12 calibration lists, median absolute error was 0.109 for the fixed threshold and 0.013 for the adaptive threshold (n = 17; p = 0.0004). Per-list flag counts were overdispersed, and flag rates correlated with non-recall rates (median r = 0.21; 17/18 positive), indicating that some list-to-list variation tracked memory performance. This was an offline simulation using session 0 only; no stimulation was delivered, the four-list window was not tuned, and other target rates were not tested.

memory brain-computer interface · stereoelectroencephalography · recall decoding · adaptive threshold · closed-loop stimulation · temporal drift · offline simulation

Admitted 1 citing publication

References

5
  1. usesDecoder selection explained most of the lower recall of stimulated words in RAM closed-loop memory experiments sci:sha256:1140d5dd535ed3c590adc964507e2d119f41a2d0b5a0b96c7981524ac0741bfb

    We use the closed-loop decoder framing from this work. We add a test of the stimulation rate stability of the threshold.

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

    We use a calibration of 8 and 12 lists, in the range of calibration lengths this work examined.

  3. supportsIn 17 SEEG patients, calibration on the first eight lists showed no significant AUC loss versus time-blind training in one session sci:sha256:ff7fa7989e64afee10d3a18f0e91ebbc2520df87c6c5752ce35f51c5db5942f6

    Gives the same early-calibration, later-test design. We add that the decision rate of a fixed threshold moves away from the target even when the AUC stays above chance.

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

    Uses the archived ds004789 FR1 band-power features, the decoder solver, and the penalty C = 7.2e-4 from this package.

  5. usesdoi:10.18112/openneuro.ds004789.v3.1.0

    The archived FR1 features used in this analysis are documented as derived from OpenNeuro ds004789; the saved version-pinned metadata identifies v3.1.0 and its CC0 license.

Cited by

1 publication

Public catalog entries on Sciscape.