Imaging-Anchored Decoder Calibration with Applicability Checks: An Exploratory Evaluation of Subject-Specific Channel Selection Across Three Public Datasets

This preprint presents an exploratory evaluation of BNS Atlas, a decoder-independent framework that uses subject-specific neuroimaging and electrode geometry to guide neural decoder calibration. Atlas checks task-signal detectability, electrode coverage, registration quality, and imaging-anchor quality, withholding guidance when applicability checks fail. Across three public OpenNeuro datasets, imaging-guided subsets of approximately 40% of recording channels were compared with 1,000 equal-size random subsets per subject. In the auditory intracranial EEG development dataset (ds003688), five of ten subjects passed all checks. Imaging-guided selection exceeded the random-subset mean in all five, with a median balanced-accuracy gain of 7.4 percentage points and individually significant gains in two subjects. Anatomical specificity remained uncertain. No consistent advantage was found in visual MEG (ds000117) or visual ECoG (ds004194). Post hoc comparisons identified examples of comparable mean decoding performance with fewer decoder-training windows, but did not establish statistical equivalence or reduced total calibration time. Coverage-based refusals avoided a loss in one subject and missed potential gains in two. Standalone verification code reproduced all 26 imaging anchors, imaging-guided and full-array scores for 11 subjects, and all 1,000 random-subset scores for two subjects. These exploratory findings motivate preregistered evaluation of when imaging guidance improves decoding and when applicability checks should withhold it.