FACT-Net: A Frequency Adapter CNN With Temporal-Periodicity Inception for Fast and Accurate MI-EEG Decoding

Abstract

Motor imagery brain-computer interface (MI-BCI) based on non-invasive electroencephalogram (EEG) signals is a typical paradigm of BCI. However, existing decoding methods face significant challenges in terms of signal decoding accuracy, real-time processing, and deployment. To overcome these challenges, we propose FACT-Net, an innovative deep-learning network for the fast and accurate decoding of MI-EEG signals. FACT-Net incorporates a Frequency Adapter (FA) module designed for processing the frequency features of MI-EEG data, as well as a Temporal-Periodicity Inception (TPI) module specifically for handling global periodic signals in MI. To evaluate the proposed model, we conduct the experiments on the cross-day dataset collected from 67 subjects and the BCIC-IV-2a dataset. The FACT-Net achieved an accuracy of 48.32% and 80.67% higher than the state-of-the-art (SOTA) approaches, demonstrating excellent performance in MI decoding. Additionally, it exhibits exceptional memory efficiency and inference time, indicating significant potential for practical applications. We anticipate that FACT-Net will set a new baseline for MI-EEG decoding. The code is available in https://github.com/Ktn1ga/EEG_FACT.

Keywords

• Brain-computer interface (BCI)
• motor-imagery (MI)
• convolutional neural networks (CNN)
• electroencephalogram (EEG)
• frequency domain analysis