In four clinically implanted patients, SEEG was less accurate than ECoG and adding SEEG did not significantly improve six-category visual decoding

We compared SEEG depth contacts with ECoG surface contacts in the same patients during six-category visual decoding. The analysis used OpenNeuro ds004194 task sixcatloctemporal in p11–p14 (2,304 balanced trials; chance accuracy 0.167), a fixed pipeline with common average reference (CAR) within each contact type, 70–150 Hz log power in two time windows, shrinkage LDA, and leave-one-run-out cross-validation. SEEG accuracies were 0.179, 0.149, 0.250, and 0.233. SEEG exceeded chance only in p13 and p14 (permutation p = 0.0099 and 0.020). ECoG accuracies were 0.672, 0.356, 0.365, and 0.299 (all p = 0.0099). SEEG accuracy was lower than ECoG in all four patients; paired tests were significant in three (p14: p = 0.077). Among 30 random ECoG subsets matched to the SEEG contact count, accuracy equaled or exceeded SEEG in 80–100% of draws. Adding SEEG to ECoG did not significantly increase accuracy in any patient; accuracy decreased by 0.022 in p11 (p = 0.021). In an exploratory analysis, 24 visually responsive SEEG contacts in p14 decoded less well than 9 responsive ECoG contacts (0.222 vs 0.444); contact selection used all trials, including test trials. Contact locations were chosen clinically, only four patients were analyzed, no anatomical analysis or bipolar reference was used, and no correction was made for multiple comparisons across patients. The data files came from the unversioned public S3 mirror; the archived provenance preserves file hashes and URLs, but does not establish an exact release match.

SEEG · ECoG · visual decoding · electrocorticography · stereo-EEG · brain-computer interface · high-frequency broadband · OpenNeuro