In 18 SEEG patients, 1- and 2-bit codes were non-inferior to full precision for offline word-recall decoding; full channel coverage improved equal-budget AUC

In per-word band-power features from 18 patients in OpenNeuro ds004789 (FR1, session 0), an offline leave-one-list-out L2 logistic decoder with training-fold quantile coding performed non-inferiorly to a full-precision z-score decoder at a 0.02 AUC margin. Two-bit coding (four levels) had median AUC 0.669 versus 0.667 for full precision (median paired change +0.008; 12/18 patients higher; one-sided exact Wilcoxon p = 3.8e-6). One-bit coding was also non-inferior (median change +0.007; Holm-adjusted p = 2.7e-4), as was two-bit coding for AUC within serial position (Holm-adjusted p = 5.7e-5). At an equal median budget of 900 bits per word, one bit on all channels outperformed eight bits on one-eighth of randomly selected channels (median AUC difference +0.031; 16/18 patients; Holm-adjusted p = 1.5e-4). Descriptive, uncorrected comparisons found small AUC increases for 2-, 3-, 4-, and 8-bit quantile codes; an exploratory clipped z-score control explained part, but not all, of these increases. This was one dataset, task, and session, with an offline fixed decoder; the analysis quantized final per-word features, not raw signals, and did not test feature drift between sessions. The pre-registration, source features and checksums, code, per-patient results, statistics, and figure are included in the supporting archive.

SEEG · word recall · neural decoding · quantization · band power · non-inferiority · channel coverage