In 18 SEEG patients, clipping band-power features at 3 SD slightly increased recall AUC, while rejecting 5% outlier training words did not

We tested two outlier controls for a within-patient SEEG word-recall decoder using archived band-power features from 18 patients in OpenNeuro ds004789 FR1, session 0. The L2 logistic-regression decoder used eight frequencies from 0–1.6 s after word onset and leave-one-list-out tests, with normalization statistics estimated from each training fold. Clipping each feature at 3 SD increased median overall AUC by 0.0026 (95% bootstrap CI 0.0007 to 0.0047; 14 of 18 patients; Holm-adjusted exact Wilcoxon p = 0.023). Within serial position, the change was +0.0016 (95% CI −0.0043 to +0.0053; p = 0.85). Rejecting the 5% of training words with the most extreme features did not increase AUC (median change −0.0041, 95% CI −0.0086 to +0.0027; p = 0.57). Extreme values made up a median 0.58% of features (Gaussian expectation: 0.27%). The most extreme words had a median recall rate of 0.13 compared with 0.22 for all words, but the outlier score alone did not predict recall (median AUC 0.475; p = 0.054, not significant). These results indicate that this L2 decoder was robust to the outliers in these features and that trial rejection was not necessary for offline recall decoding in this dataset. The analysis covered one offline session per patient; only one clipping threshold and one rejection rate were tested, and the results do not apply to raw-signal artifacts.

SEEG · word recall · memory decoding · brain-computer interface · band power · outlier clipping · trial rejection · logistic regression · OpenNeuro ds004789