A cross-patient SEEG decoder did not predict recall within serial position in 7 held-out RAM FR1 patients; patient-specific decoders did better

Question: can a recall decoder for a closed-loop memory BCI work in a new SEEG patient without labeled calibration data? Data: archived per-word band-power features (3–140 Hz, 0–1.6 s after word onset) of 7 SEEG patients in OpenNeuro ds004789 (RAM FR1, session 0, 500 Hz). Method: each patient's channels were z-scored without labels and averaged, which gave 8 channel-agnostic features per word. A logistic decoder was trained on 6 patients and tested on the held-out patient. The primary endpoint was the AUC from recalled/not-recalled pairs with the same serial position. The plan was fixed before the analysis. Result: the primary test failed. Median within-serial-position AUC was 0.487 (3/7 patients above 0.5; one-sided exact Wilcoxon p = 0.47). Only 2 patients were above chance by permutation. Causal z-scoring (first 2 lists) gave the same result. Patient-specific all-channel decoders gave median 0.611 (6/7 above 0.5) with leave-one-list-out. On a fixed test set (last 8 lists), they gave 0.64–0.72 with only 1–2 labeled lists, versus 0.52 for the cross-patient decoder. The cross-patient weights had consistent signs at 5 Hz (negative), 80 Hz (positive) and 140 Hz (negative), but this shared pattern did not rank words in new patients. Limits: n = 7, one session, a convenience subset, and only one type of channel-agnostic feature. The package has code, results, a figure and reproduction steps.

SEEG · free recall · memory decoding · cross-patient generalization · calibration · serial position · band power · AUC · RAM FR1