Cross-Channel Prediction, Classification, and Individual Retention in Low-Channel EEG Adaptation

Research question

With fewer EEG channels, adaptation can target feature prediction or task classification. This work tests whether gains transfer between these objectives and whether improved future performance coexists with retained performance on an earlier test for each participant.

Study design

Four public EEG datasets provide comparisons with fixed information-release rules and matched controls. Initial test data remain permanently excluded from fitting. The analysis separates paired cross-channel calibration, task-label updating, and constraints intended to retain earlier performance.

Main findings

• In the 42-participant Lee confirmation group, paired calibration improved prediction of omitted-channel features. For a fixed downstream classifier, the AUC difference between updated and frozen mappings was −0.00220, with a 95% interval of −0.00968 to +0.00402. A classification gain was unconfirmed.
• In the 62-participant Stieger development cohort, shared updating with a 25% new-label budget improved mean future balanced accuracy by 2.479 percentage points relative to zero new labels. The 95% interval, conditional on the evaluated setup, was 1.930–2.982.
• 19 participants scored lower on the permanent old test than their own initialized model; 12 declined by at least five percentage points. All 19 had positive future gains against the zero-label control.
• The tested training-risk constraints passed the future-performance tolerance but failed the joint criterion requiring reduced old-test harm.

Interpretation

Cross-channel prediction, downstream classification, and individual retention require separate measurements. The future-gain comparison uses a zero-new-label control; the retention comparison uses each participant’s initialized model. Reporting both makes the benefits and costs of updating visible at the participant level.

Research extension

Sparse EEG for Movement and Motor Learning extends this work toward movement measurement and motor-learning feedback. The proposed direction is at the research-planning stage, with a concept demonstration using synthetic signals.