SEEG encoding band power did not predict spatial-memory accuracy in 7 patients

We used SEEG band power during encoding to predict later spatial-memory accuracy. We used OpenNeuro ds005522 (RAM YC1 spatial navigation, snapshot 1.0.0, session 0, bipolar montage). We included 7 patients with only SEEG channels and 48 object trials each. For each object, the patient drove to it two times (learning). Then the patient tried to find the location from memory (test). The label was a median split of the percentile error, which corrects for object position. Features were log power in 5 bands (3–150 Hz) for each bipolar channel, from the first 3 s of each learning drive. A shrinkage LDA decoder used 5-fold cross-validation grouped by object trial. Result: the mean AUC was 0.522 (SD 0.072). The exact sign-flip test gave p = 0.44. The AUC was above 0.5 in 4 of 7 patients. No single band decoded above chance (mean AUC 0.47 to 0.54, all p >= 0.22). The univariate correlation between power and error was near 0 in all bands (all p >= 0.11). Thus a simple whole-brain spectral readout of the encoding state is not a reliable memory-BCI signal for this task when there are about 48 trials per patient. Limits: only large effects can be detected, and region selection and other time windows were not tested. The package contains the code (byte-range EDF reader, extraction, analysis), features, S3 keys, sha256 hashes and the reproduction steps.

SEEG · spatial memory · memory decoding · band power · brain-computer interface · cross-validation · OpenNeuro · YC1