In 21 RepFR1 patients, events.tsv sample was word-locked and onset was not

A memory brain-computer interface must determine during word encoding whether a patient will later recall a word. Three library studies of RepFR1 SEEG data from OpenNeuro ds005411 found no such recall read-out; one reported mean held-out AUC 0.498 in 21 patients. We tested whether the events.tsv sample column (EDF sample index) or BIDS onset column (seconds) was locked to words. Before reading neural data, we wrote and hashed a plan. In 21 patients, we compared 70–150 Hz power at both candidate times for 81 words per patient, using the first three lists with a recall. The sample time had more responsive channels in 16 patients, onset in none, and five tied (mean fraction 4.04% vs 0.41%; exact two-sided sign-flip p = 3.1e-5). Median |t| was larger at sample in 18 patients (mean 0.88 vs 0.73; p = 7.1e-4). The response at sample peaked 0.3–0.4 s after the word; onset had no word-locked response. Thus sample was word-locked, and the archived features used by study 02694702 had valid timing. The plan therefore did not call for re-extracting features at onset, and this was not done. In 9 patients the columns also ran at different rates (4.5–52 ppm), so a constant onset correction would not suffice; the difference changed by up to 0.44 s over lists 0–9. The study tested only three lists per patient, resolves alignment to about 0.1–0.2 s, included all bipolar channels (including ECoG-typed channels), and did not test new recall decoders. The archived source files do not establish an exact OpenNeuro release; the README preserves the source URL and file hashes.

RepFR1 · SEEG · BIDS · event timing · word encoding · high gamma · OpenNeuro · memory · electrophysiology