FR1 decoding predicted PAL1 recall in seven RAM patients, with significance dependent on analysis rules

In an offline analysis of session 0 data from seven SEEG patients who completed RAM free recall (FR1) and cued word-pair recall (PAL1), a per-patient L2 logistic decoder trained on FR1 channel-by-frequency features predicted PAL1 recall without PAL1 labels (mean AUC 0.595; median 0.576; 6/7 above 0.5; one-sided exact Wilcoxon p = 0.023). The per-patient results reproduced a prior SCI analysis exactly. The mean AUC remained 0.59–0.60 across four combinations of PAL1 label and inclusion rules, while p values ranged from 0.023 to 0.078. A channel-mean model did not transfer (mean AUC 0.520, p = 0.29), and a fixed spectral-tilt measure did not predict PAL1 recall (AUC 0.464, p = 0.89); the decoder exceeded the tilt in all seven patients (p = 0.016). Effect-map results suggested channel-specific shared information despite opposite mean frequency profiles. Reverse transfer had mean AUC 0.575 (p = 0.039). These data support a small, channel-specific transfer effect in this cohort, but the sample is small, significance is sensitive to reasonable analysis choices, and the work is not an independent replication or a closed-loop test.

SEEG · memory encoding · free recall · cued recall · transfer learning · logistic regression · brain-computer interface · AUC