Benchmarking Label-Revealed Online Updates for EEG BCI Decoding

Bogdan Kozyrskiy; Artem Grachev; Abraham I. Camelo Guerrero

cs.LGOct 5, 2026

Benchmarking Label-Revealed Online Updates for EEG BCI Decoding

Authors: Bogdan Kozyrskiy, Artem Grachev, Abraham I. Camelo Guerrero

Organizations: Independent Researcher, France · Independent Researcher, Germany · School of Information Technology, York University, Toronto, Canada

Abstract

Electroencephalography (EEG) signals drift over time, which can cause static brain-computer interface (BCI) models to degrade in practice. We present a benchmark for online adaptation and compare two widely used pipeline families, Common Spatial Patterns (CSP) and Riemannian covariance-based methods, under time-ordered prequential (test-then-train) evaluation. We examine (i) which pipelines benefit most from label-revealed updates, (ii) whether controlled forgetting of older data improves robustness, and (iii) how a minimal-calibration cold start compares with starting from a pretrained model. Across four datasets (three motor-imagery datasets and one movement-decoding dataset), label-revealed online updates improve 13 of 14 model/dataset pairs on the two largest streams, with relative accuracy gains of up to about 18% over a frozen model. A Shapley-based data-valuation analysis over temporal blocks assigns the largest mean value to the most recent block in each of the three analyzed datasets, while older blocks retain positive value.

Figures & tables

Figure 1 : Expanding-window (walk-forward) evaluation scheme.

Figure 2 : Rolling-window (sliding-window) evaluation scheme.

Table 3

Figure 3 : Prequential accuracy on BCIC 2a for ACM+TS+SVM, frozen vs. online refit (9 subjects; bold curves are the across-subject mean, shaded bands the standard error, thin curves individual subjects). Left: trailing-window accuracy $A_{k}$ (Eq. 2 , $w=25$ ). Right: cumulative accuracy $C_{k}$ (Eq. 1 ). Online refit remains above frozen for most of the stream and the advantage grows over time.

	ACM+TS+SVM		FBCSP
Metric	Frozen	Online refit	Frozen	Online refit
Final $C_{k}$	0.689	0.767	0.639	0.673
Final $A_{k}$	0.644	0.760	0.627	0.649

Table 4 : Final prequential accuracy on BCIC 2a (9 subjects, 288-step stream), frozen vs. online refit. Holm-adjusted paired Wilcoxon tests give

p=0.059

and

0.046

for ACM+TS+SVM final

C_{k}

and

A_{k}

, and

p=0.258

and

0.499

for FBCSP.

Panel of Figure 4 — Figure 4 : Training-sample contribution analysis on BCIC. Left: paired tests comparing summed Shapley values across training blocks. Right: influence profile as a function of temporal distance.

Figure 5 : Exact block-level (part) Shapley value on BCIC 2a. The 432-trial training pool is split into $P=6$ contiguous temporal blocks (oldest to newest), each treated as a single Shapley player, so the value is computed by exact enumeration of all $2^{6}$ coalitions. Bars are the across-run mean $\pm$ standard error for FBCSP and ACM+TS+SVM. The newest blocks receive the largest credit in both pipelines.

Panel of Figure 6 — Figure 6 : Cross-dataset comparison of the exact block-level Shapley valuation. The newest temporal block carries the largest mean value in both pipelines and on both datasets; the gradient is flatter on the small ALEXMI stream and sharper on the larger Schirrmeister 2017 streams.

Figure 7 : Smoothed signed helpful influence versus train–test temporal distance for a single subject (subject 1, seed 0). Values above zero are helpful, below zero harmful. FBCSP is slightly positive at short distances and turns negative in the far tail, whereas ACM+TS+SVM stays much closer to zero across the axis, indicating that FBCSP is more sensitive to temporal distance for this subject.

Explore similar work

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks
Jun 1, 2026Yangxuan Zhou, Sha Zhao, Jiquan Wang +2Brain-Computer Interfaces EEG Decoding
State Key Laboratory of Brain-machine Intelligence and College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China · MOE Frontier Science Center for Brain Science and Brain-Machine Integration, Zhejiang University, Hangzhou, Zhejiang, China
Separating personal from population gains when calibrating EEG foundation models for new users
Sep 28, 2026Xilin Tao, Kani ChenCross-Subject EEG Decoding MI Classification
Department of Mathematics, The Hong Kong University of Science and Technology, Hong Kong SAR, China · Department of Industrial Engineering and Decision Analytics, The Hong Kong University of Science and Technology, Hong Kong SAR, China
Same Brain, Different Prediction: How Preprocessing Choices Undermine EEG Decoding Reliability
May 8, 2026Dengzhe Hou, Zihao Wu, Lingyu Jiang +3Uncertainty Quantification EEG Decoding
Tohoku University · University of Georgia · Texas A&M University +1

Jun 1, 2026, cs.AIJ/K move · Enter open · S save

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks

Yangxuan Zhou, Sha Zhao, Jiquan Wang, Shijian Li+1

State Key Laboratory of Brain-machine Intelligence and College of Computer Science and Technology, Zhejiang University, Hangzhou, Zhejiang, China · MOE Frontier Science Center for Brain Science and Brain-Machine Integration, Zhejiang University, Hangzhou, Zhejiang, China

Brain-Computer Interfaces EEG Decoding Continual Learning Replay-Based Continual Learning EEG Foundation Models

Open paper AskPDF arXiv

Electroencephalography (EEG) is the cornerstone of non-invasive brain-computer interfaces (BCIs), yet conventional decoding relies on fragmented, task-specific architectures that severely limit cross-task scalability. While EEG foundation models pre-trained on massive corpora promise universal brain decoding, current post-training depends on task-isolated fine-tuning. This static paradigm restricts knowledge transfer across heterogeneous tasks, hinders model scalability, and incurs computational and storage overheads that scale linearly with task count. To overcome these bottlenecks, we formulate downstream adaptation as a cross-task continual learning problem and propose EvoBrain, a dynamic, task-aware continual learning framework for unified EEG decoding. EvoBrain addresses the plasticity-stability trade-off via two complementary components: (1) Neuro-Spectral Task Normalization (NSN) aligns incoming tasks with historical statistics while recalibrating spectral responses to handle distributional and neuro-spectral shifts; and (2) Response-Affinity Distillation (RAD), combined with time-dependent replay, preserves old-task response geometry and promotes selective knowledge transfer between spectrally compatible tasks, effectively mitigating forgetting. Extensive evaluations across six distinct BCI tasks demonstrate that EvoBrain consistently surpasses state-of-the-art methods across diverse foundation backbones, optimally balancing plasticity and stability. To our knowledge, this work pioneers cross-task continual learning in the EEG domain, advancing the realization of a unified, one-for-all brain decoding system.

Abstract

Explore similar work

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks

Separating personal from population gains when calibrating EEG foundation models for new users

Same Brain, Different Prediction: How Preprocessing Choices Undermine EEG Decoding Reliability