A Confidence-Gated Source Selection Strategy for Cross-Session Transfer in Brain-Computer Interfaces

Published · Frontiers in Human Neuroscience - Brain-Computer Interfaces

Recommended citation: Shen Y and Degras D (2026) A confidence-gated source selection strategy for cross-session transfer in brain–computer interfaces. Front. Hum. Neurosci. 20:1895016. doi:10.3389/fnhum.2026.1895016. https://doi.org/10.3389/fnhum.2026.1895016

Published: 28 August 2026 in Frontiers in Human Neuroscience, 20:1895016. DOI: 10.3389/fnhum.2026.1895016.

Overview

Historical EEG sessions are not automatically interchangeable sources for a new recording session. This study evaluates how source sessions should be pooled, weighted, or selected for within-subject, cross-session motor-imagery EEG transfer.

The comparison uses matched feature, classifier, domain-adaptation, and distance-method banks so that differences between source-utilization strategies are not confounded with unequal model searches. Experiments cover Stieger2021 (62 participants) and BNCI2014-004, with final-test labels kept out of configuration selection.

Strategies compared

  • MAP - Merge and Adapt: pool all available source sessions before adaptation.
  • DWP - Distance-Weighted Pooling: weight source sessions by their measured proximity to the target session.
  • MMP - Minimum-Distance Multi-source: retain source sessions through confidence-interval-gated distance comparisons.
  • BDP - Bridge-Domain Proxy: use bridge and far sessions to separate proxy configuration selection from final target-session evaluation.

Main findings

  • On Stieger2021, MAP, DWP, and BDP were statistically tied as the top pipelines after data-driven configuration selection.
  • BDP remained in the MAP/DWP selected-performance range while its shared confidence-interval gate retained 74.29% of source sessions.
  • The BDP pipeline used 37% of MAP’s complete benchmark-search runtime on Stieger2021.
  • These results support an efficiency interpretation for selective source use; they do not establish a uniform accuracy improvement over full pooling.

Evaluation controls

  • Leave-one-session-out evaluation with target labels held out until final testing.
  • Matched configuration banks across source-utilization strategies.
  • Auditable train, bridge, far, and target-session role assignments.
  • Separate reporting of selected-configuration performance, source utilization, and complete benchmark-search runtime.

Companion software

CrossDA provides the CLI and Python API used to run MAP, DWP, MMP, and BDP experiments. It writes per-subject summaries, detailed run records, and session-role tables for downstream audit and analysis.

View CrossDA on GitHub

Read the published article