Coarse MNI-region features did not improve a cross-patient SEEG word-recall decoder in 18 patients

We tested whether coarse anatomical location adds transferable information to a cross-patient decoder of later word recall from stereo-EEG (SEEG). Using band-power features for 18 patients from the FR1 task, session 0, we compared an eight-feature channel-mean model with models that added region-by-frequency means from six or 12 coordinate-defined regions. In leave-one-patient-out evaluation, median AUCs were 0.556 for the channel-mean model, 0.553 for the six-region model, and 0.554 for the 12-region model. The primary paired comparison (six-region minus channel-mean) had median difference -0.007; 6 of 18 patients had higher AUC, and the two-sided exact Wilcoxon p value was 0.52. The secondary 12-region comparison had median difference -0.004; 8 of 18 patients had higher AUC (p = 0.37). All three models were above chance. These results show no average improvement from these coarse regions in this sample. The study used one dataset and task, session 0 only, offline AUC, and 18 patients. Regions were coordinate boxes rather than atlas labels; missing regions were set to zero, and the test had low power to detect small gains. More patients or stronger regularization may allow location to help.

stereo-EEG · word recall · cross-patient decoding · band power · anatomical regions · leave-one-patient-out · AUC · OpenNeuro