Toward Adaptive BCIs: Enhancing Decoding Stability via User State-Aware EEG Filtering

TL;DR

The paper tackles BCI decoding instability caused by fluctuating user states by introducing a state-aware EEG filtering framework that estimates attention from EEG and prunes unreliable epochs before decoding. It implements a dual-stage approach: (i) attention-based masking using the Alpha–Theta ratio $ATr_i = rac{E_{ ext{α}}(i)}{E_{ ext{θ}}(i)}$ to filter distracted segments, and (ii) confidence-guided, loss-based sample weighting with $L_i = \mathcal{L}(y_i, f(x_i; \theta))$ and $w_i = 1$ if $L_i < \lambda$ (increasing $\,\lambda$ during training). Experiments on MOABB motor-imagery benchmarks show significant accuracy gains across backbones (EEGNet, ShallowConvNet, DeepConvNet) and datasets (BNCI2014004, Zhou2016), along with reduced inter-subject variability, validating the neurophysiological relevance of the filtering. By enabling robust learning from attention-consistent neural representations, the approach highlights the practical potential of leveraging brain-derived state information for adaptive, long-term reliable EEG-based BCIs.

Abstract

Brain-computer interfaces (BCIs) often suffer from limited robustness and poor long-term adaptability. Model performance rapidly degrades when user attention fluctuates, brain states shift over time, or irregular artifacts appear during interaction. To mitigate these issues, we introduce a user state-aware electroencephalogram (EEG) filtering framework that refines neural representations before decoding user intentions. The proposed method continuously estimates the user's cognitive state (e.g., focus or distraction) from EEG features and filters unreliable segments by applying adaptive weighting based on the estimated attention level. This filtering stage suppresses noisy or out-of-focus epochs, thereby reducing distributional drift and improving the consistency of subsequent decoding. Experiments on multiple EEG datasets that emulate real BCI scenarios demonstrate that the proposed state-aware filtering enhances classification accuracy and stability across different user states and sessions compared with conventional preprocessing pipelines. These findings highlight that leveraging brain-derived state information--even without additional user labels--can substantially improve the reliability of practical EEG-based BCIs.

Figures (3)

• Figure 1: Illustration of the user state-based binary masking process for robust decoding against state fluctuations. Energy from attention-related frequency bands is extracted from EEG signals and mapped to feature space to generate masks, which are then applied to the existing training dataset for refinement.
• Figure 2: Performance comparison between baseline and proposed methods across various network architectures on (a) BNCI2014004 and (b) Zhou2016 datasets. Stars denote statistical significance between baseline and our proposed method: $^*$$\textit{p} < 0.05$, $^{**}$$\textit{p} < 0.01$, $^{***}$$\textit{p} < 0.001$.
• Figure 3: Alpha PSD topographical maps demonstrate localized patterns characteristic of motor imagery task in included samples, while excluded samples exhibit elevated occipital alpha power indicative of user distraction, with all visualizations derived from the Zhou2016 dataset.