Under assumed timing, 34 of 96 single labels were wrong for 90% target load sets in 12 final SEEG sessions

Can load prediction sets retain useful output after calibration in a separate SEEG session? We tested public OpenNeuro ds004752 version 1.0.1. Twelve patients had separate training, calibration, and final test sessions. The test used 344 eligible trials. Each reader used eight selected low-band features and assumed task timing. With a 90% label coverage target, mean coverage was 90.0% (95% patient bootstrap interval, 85.9% to 93.9%). The reader gave one label on 27.7% of trials. Error among these outputs was 36.0% with equal patient weights (28.1% to 45.5%). In pooled counts, 34 of 96 single labels were wrong. The other 248 sets contained both loads. Separate calibration for each load increased coverage to 93.4%, but reduced single-label output to 18.2%. Its error among single labels remained 36.2%. A constant-prior control gave both labels on every test trial. Thus, high label coverage alone did not establish reliable single-label output for this reader. The package adds a reproducible uncertainty test for later SEEG sessions. It preserves all 68 source feature files, exact source metadata and hashes, code, fitted readers, prediction sets, three figures, and source retrieval instructions. Source timing is assumed. These are task load labels in a selected epilepsy cohort. The study did not test online control or brain stimulation.

brain-computer interface · SEEG · conformal prediction · prediction sets · working memory · label coverage · session transfer