Toward Stable Brain-Computer Interfaces: Revealing and Addressing Prediction Fluctuations in EEG-Based BCIs

Runhe Zhou, Yi Ding, Guanxiang Huang, Cuntai Guan

Keywords: Deep Learning, Brain-Computer Interfaces, Prediction Fluctuations, Model Robustness

TL;DR: This paper explores prediction fluctuations in deep learning-based brain-computer interfaces, introducing stability metrics and mitigation strategies to enhance robustness.

Abstract: Brain-Computer Interfaces (BCIs) are increasingly used in areas such as neurofeedback and mental healthcare, where reliable real-time feedback is essential. While deep learning (DL) has greatly improved Electroencephalography (EEG)-based BCIs by boosting accuracy in tasks like emotion recognition, attention detection, and workload assessment, current models often suffer from \textit{temporal instability}. Predictions fluctuate erratically across consecutive windows, contradicting the slow-changing nature of cognitive states and producing inconsistent feedback that undermines user engagement. Existing metrics and post-processing methods fail to capture or resolve this issue effectively. We address this gap through three contributions: (1) a systematic study of prediction fluctuations across datasets, tasks, and representative models; (2) two new stability metrics, Frequency-weighted Spectral Entropy (FSE) and First-Order Difference Standard Deviation (FDS), that directly measure temporal irregularities; and (3) TRin (Temporal Robustness integrated BCI), a fluctuation-aware training framework combining stability-driven losses with curriculum learning. Experiments on three public datasets show that TRin consistently reduces fluctuations while improving accuracy. By introducing stability as a core evaluation dimension, this work provide a new way for more robust and effective real-time BCIs.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 23181