FR1 pre-training did not significantly improve PAL1 SEEG recall decoding in seven patients

We assessed whether a patient's free-recall (FR1) session reduced the number of labelled paired-associate (PAL1) lists needed to calibrate a within-patient SEEG decoder. Using archived features from seven RAM patients, we compared PAL1-only logistic regression, FR1-only, pooled FR1 and PAL1 training, and PAL1 training with an FR1-centred L2 penalty. Features were log10 Morlet power at eight frequencies (3–140 Hz), averaged over 0–1.6 s after item onset on bipolar channels shared across sessions. We trained on the earliest k = 0, 1, 2, 3, 4, 6, 8, or 10 PAL1 lists and evaluated all models on the same later lists. For the prespecified primary comparison, the median per-patient AUC gain for pooled training over PAL1-only training averaged across k = 1, 2, and 3 was +0.008 (one-sided exact Wilcoxon W+ = 17, p = 0.34; n = 7), not statistically significant. Randomly selecting calibration lists gave a median gain of +0.067 (p = 0.11), also not significant. The FR1-only model had median AUC 0.604; at k = 10, its difference from PAL1-only was −0.009 (p = 0.81). Pooled performance was nearly the same as FR1-only. These results do not establish that pre-training reduces PAL1 calibration needs. Interpretation is limited by seven patients, small test sets (including 5–7 recalled pairs for two patients), one regularization value, and session-wise z-scoring that used unlabelled test-session data. The analysis plan was written before the run. Source features were archived from earlier SCI contributions; their dataset provenance and file-level hashes are retained in the accompanying source files.

SEEG · memory decoding · memory BCI · free recall · paired-associate learning · transfer learning · calibration · AUC · intracranial EEG