Nine archived SEEG recall decoders show moderate patient ranking agreement and substantial sampling noise

We asked whether per-patient word-recall decoding AUC is a stable patient property in OpenNeuro ds004789 (RAM FR1, SEEG). We listed files from 108 SCI publications that name ds004789 or FR1 SEEG and copied 81 per-patient tables. From 14 tables, we selected the main within-patient all-feature recall AUC column using headers only, then removed five duplicate AUC vectors, leaving nine pipelines. Across 18 patients, median pairwise Spearman rho was 0.69 (IQR 0.45–0.78; minimum 0.21; 28 pairs); the 500-permutation null median was 0.00 (95th percentile 0.09; p = 0.002). Kendall W was 0.62 among the eight patients shared by all pipelines. In an additive variance decomposition, patient, pipeline, and residual components were 67%, 8%, and 25%; single-pipeline ICC was 0.70. Mean AUC was 0.64 and patient SD was 0.046. Hanley–McNeil sampling SEs from 40–74 recalled words per session ranged from 0.032 to 0.062 (mean 0.045); estimated sampling noise was 47% of variance between patient means. Sixteen of 18 patients were above chance by more than 1.96 SE. Mean AUC had little association with word count (rho 0.03), channel count (0.22), or recall rate (−0.18). Compared with archived ds004752 working-memory load results (rho 0.86, W 0.97, patient SD 0.15), recall decodability differed less between patients. One session cannot reliably rank FR1 patients for a recall BCI; ranking needs a second independent session or more recalled words. Limits: pipelines are not independent and often reuse band-power features; this is one cohort and one offline session; values are as archived and source code was not rerun; only eight patients occur in all pipelines; Hanley–McNeil assumes independent words; and column selection reflects the authors’ judgment. The exact ds004789 release underlying the archived analyses is not established.

intracranial EEG · SEEG · word recall · memory decoding · brain-computer interface · area under the curve · pipeline stability · sampling noise