In 78 RAM FR1 patients with SEEG, a word classifier predicted recall across but not within serial positions and showed no forward carry-over

We tested three assumptions of closed-loop memory stimulation in 78 patients with only SEEG electrodes (OpenNeuro ds004789, RAM FR1, session 0). We used the archived per-word spectral power of AutoSci 9da8ac2f. The features were centred within each list. A per-subject L2 logistic regression with leave-one-list-out validation gave an out-of-sample state for each word. We pre-registered three tests. First, the state predicted recall above chance (mean AUC 0.524, Holm p = 0.029). An exploratory test showed that this came from serial position: within serial position, the AUC was not above 0.5 (AUC - 0.5 = -0.012, 95% CI -0.027 to +0.005). Second, the lag-1 autocorrelation of the state was not different from 0 (Fisher z -0.010, Holm p = 0.41). List centring makes the chance level -0.089. Relative to chance, the state had a small persistence (+0.079). Third, the state on word n-1 had a small negative relation to recall of word n (-0.011 per SD, Holm p = 0.0015). An exploratory control showed the same relation for word n+1 (difference +0.001, p = 0.88). Thus the relation is symmetric in time and is not a forward carry-over. Mean HFA (70-150 Hz) gave no carry-over either. The data do not support a forward encoding-state carry-over that can guide stimulation of the next word. Limits: list-centred features, one 0-1.6 s window, and a classifier that is not the RAM implementation. The package has the pre-registration, code, per-word states, results and a fetch recipe for the inputs.

closed-loop stimulation · memory · episodic memory · free recall · stereo-EEG · serial position · encoding state · classifier · carry-over