Confidence-Calibrated Source Screening for Cross-Session Motor-Imagery EEG under Distribution Shift
Working Paper
Research question
Historical EEG sessions differ in their similarity to a new recording. A ranking based on estimated distances does not establish which differences exceed sampling uncertainty. This work asks when the evidence supports excluding a historical session from the training pool.
Method
The method compares session covariance summaries using unlabeled target trials. It excludes a source only when simultaneous confidence bounds identify another source with a smaller population discrepancy from the target. Each bootstrap replicate reuses one target resample across all source comparisons, preserving their shared-target dependence.
Under the stated regularity conditions, the retained set contains all population discrepancy minimizers at the prescribed asymptotic confidence level. This guarantee concerns distributional discrepancy; downstream decoding is evaluated separately.
Empirical findings
- Four public longitudinal EEG cohorts were evaluated, including one exploratory cohort, with decoding measured on held-out target trials.
- Screening retained 48%–71% of candidate sessions. Mean balanced accuracy differed from all-source pooling by −0.5 to +2.6 percentage points.
- Three cohorts met a two-percentage-point non-inferiority criterion; the fourth remained inconclusive.
- In a simulation isolating shared-target dependence, contrast coverage was approximately 95% with shared-target resampling and 85% with independent target resampling.
The output combines a reduced training pool with the comparisons and confidence bounds supporting each exclusion.
Relationship to the published source-selection study
The published cross-session transfer study compares pooling, weighting, and operational confidence-gated selection within transfer pipelines. This manuscript develops simultaneous calibration and set-valued screening for comparisons that share the same target sample. The two studies reuse Stieger2021 and BNCI2014_004, with different protocols and endpoints.