Ridge decoding did not predict later recall from encoding power in 81 pure-SEEG subjects

In 81 pure-SEEG subjects from the RAM FR1 task (OpenNeuro ds004789, version 3.1.0; session 0), a leave-one-list-out ridge classifier did not predict later word recall from encoding power across five bands and all bipolar pairs better than a serial-position-preserving permutation null. The planned score with a fold intercept had mean AUC 0.456 versus null 0.457 (effect -0.001, 95% CI -0.019 to +0.017, sign-flip p = 0.89). A post-hoc score without the fold intercept had AUC 0.504 versus null 0.496 (effect +0.008, 95% CI -0.008 to +0.024, p = 0.35). Secondary models were not significant after Holm correction; theta-only power had raw p = 0.040 and Holm p = 0.12. Serial position alone gave AUC 0.561. The upper confidence bound for the no-intercept effect was about 0.024 AUC. This result applies to five broad periodogram bands, session 0, clinical electrode coverage, and one linear model; it does not exclude nonlinear or time-resolved information.

subsequent memory effect · free recall · stereo-EEG · ridge regression · decoding · periodogram power · serial position · OpenNeuro ds004789