Summed per-shaft SEEG scores were non-inferior for word-recall decoding, with a small AUC loss (ds004789 FR1, 18 patients)

We tested whether a modular SEEG decoder, with one L2-logistic model per shaft and a sum of shaft logits, was non-inferior to one joint model on all bipolar pairs for predicting later recall of a word from band power during encoding. The analysis used archived per-word features from OpenNeuro ds004789 v3.1.0 (FR1, session 0), with 18 patients and list-grouped five-fold cross-validation. Median AUC was 0.666 for the sum and 0.670 for the joint model. The sum met the 0.02 AUC non-inferiority margin (mean difference -0.009, 95% bootstrap CI [-0.016, -0.001], exact one-sided sign-flip p = 0.0066), while also performing slightly worse in the two-sided comparison (p = 0.045). The result was similar when comparing words at the same list position (mean difference -0.006 [-0.017, 0.004], p = 0.010). The sum sent a median of 12.5 values per word, compared with 900 for the joint decoder, an 80-fold reduction. A learned stacker and the best single shaft performed worse than the simple sum. These results suggest that recall information was distributed across shafts and that cross-shaft covariance carried a small amount of information. The analysis was offline, used one session per patient and one band-power feature set, and did not tune regularization per shaft. A median AUC near 0.67 indicates a weak decoder; conclusions apply to this signal level and setup.

stereo-EEG · word recall · memory decoding · band power · modular decoder · non-inferiority · AUC · cross-validation