The encoding decoder score predicted current word recall but not next-word recall in catFR1

We tested whether a held-out encoding-period decoder score for a word predicted recall of the next word in categorized free recall (catFR1). For each of 8 patients, we fit an L2 logistic regression to archived per-word log-power features from all bipolar channels and 8 frequencies, using leave-one-list-out folds. The decoder read out recall of the current word (mean AUC 0.592; all 8 patients above 0.5; exact one-sided p = 0.0039). It did not predict recall of the next word above chance for adjacent same-category pairs (mean AUC 0.495; Holm-adjusted p = 0.77) or different-category pairs (0.454; p = 0.14). The pre-registered primary same-minus-different AUC difference was +0.041 (5/8 patients positive; exact two-sided sign-flip p = 0.094). Recall of adjacent words was more strongly coupled for same-category pairs (mean phi 0.321) than different-category pairs (0.105; difference +0.216, 6/8 positive, p = 0.047), a secondary test not corrected for multiple tests. Thus, the encoding decoder score did not carry the category-related coupling in recall. The result is limited to 8 patients, one session per patient, one fixed decoder, and no stimulation data. Same-category pairs began at odd serial positions and different-category pairs at even positions; position-mean removal does not remove position-specific decoder effects. The test had low power for differences below about 0.05 AUC.

categorized free recall · memory encoding · next-word prediction · SEEG · intracranial EEG · logistic regression · AUC · semantic categories · recall coupling · catFR1