Non-Stationarity in Brain-Computer Interfaces: An Analytical Perspective

TL;DR

This paper surveys non-stationarity and covariate shift in EEG-based BCIs, examining causes across within-session, across-session, and inter-subject variability and their impact on generalization. It reviews a broad set of mitigation strategies, including signal normalization, domain/adaptation, online learning, and XAI-driven approaches, and maps them onto datasets and evaluation practices. The authors identify persistent challenges such as data scarcity, limited cross-subject generalization, and lack of standardized benchmarks, and they propose future directions emphasizing data frameworks, hybrid algorithms, and user-centered co-adaptation. Overall, the work provides a structured, comprehensive view of how to build more robust, calibration-free BCIs by integrating preprocessing, representation learning, adaptive decoding, and practical hardware considerations.

Abstract

Non-invasive Brain-Computer Interface (BCI) systems based on electroencephalography (EEG) signals suffer from multiple obstacles to reach a wide adoption in clinical settings for communication or rehabilitation. Among these challenges, the non-stationarity of the EEG signal is a key problem as it leads to various changes in the signal. There are changes within a session, across sessions, and across individuals. Variations over time for a given individual must be carefully managed to improve the BCI performance, including its accuracy, reliability, and robustness over time. This review paper presents and discusses the causes of non-stationarity in the EEG signal, along with its consequences for BCI applications, including covariate shift. The paper reviews recent studies on covariate shift, focusing on methods for detecting and correcting this phenomenon. Signal processing and machine learning techniques can be employed to normalize the EEG signal and address the covariate shift.

Figures (10)

• Figure 1: EEG relative power topography in theta, alpha, and beta bands showing variability due to cognitive and physiological fluctuations, such as fatigue and attention. Adapted from Lin et al. Lin2020_fatigue_topography.
• Figure 2: Illustration adapted from Saha and Baumert (2019) saha2019intra, showing intra- and inter-subject variability in EEG-based sensorimotor BCI. Inter-subject variability (top) reflects differences in EEG feature distributions across individuals due to anatomical and functional differences, while intra-subject variability (bottom) captures session-to-session changes within the same individual caused by physiological or psychological fluctuations. Source: Adapted from Frontiers in Computational Neuroscience. saha2019intra
• Figure 3: An anatomical graphic of the EEG signal path and the corresponding electrical model, illustrating electrode-skin interface, conductive gel, and tissue layers. Variations in these impedances influence EEG signal quality and contribute to noise and distortion, which are key factors affecting EEG non-stationarity and classification robustness. Adapted from Seok et al. (2021) Seok2021.
• Figure 4: EEG signals perturbed by various artifacts, including power line noise, eye movements, eye blinks, and muscle movements, along with their corresponding power spectral densities (PSDs). These artifacts overlap with frequency bands of interest, complicating signal analysis. Adapted from Nicolas-Alonso and Gomez-Gil (2012) GarciaMolina2004.
• Figure 5: EEG channel configuration over the sensorimotor cortex and the MI training paradigm illustrating learning-induced adaptation in BCI systems. Initially, broad cortical activations become more localized with practice, necessitating the model's adaptive recalibration. Adapted from Rezeika et al. (2021) Rezeika2021.