Enhanced SSVEP Bionic Spelling via xLSTM-Based Deep Learning with Spatial Attention and Filter Bank Techniques

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Steady-State Visual Evoked Potentials (SSVEPs) are crucial for brain-computer interfaces (BCIs), especially for individuals with aphasia.
• Existing transformer-based methods often underutilize SSVEP frequency information and involve redundant computations.

Purpose of the Study:

• To propose a novel deep learning architecture, SED-xLSTM, for enhanced SSVEP signal analysis.
• To improve the efficiency and accuracy of SSVEP-based BCIs by integrating time and frequency domain information.

Main Methods:

• Developed a stacked encoder-decoder (SED) network architecture incorporating an xLSTM model and spatial attention mechanism (SED-xLSTM).
• Utilized low-channel spectrograms as input and applied filter bank techniques to capture harmonic information.
• Employed a gating mechanism for effective extraction and fusion of high-dimensional spatial-channel semantic features.

Main Results:

• SED-xLSTM demonstrated superior performance in classification accuracy and information transfer rate across three public datasets.
• The model outperformed existing methods, especially in cross-validation scenarios across different temporal scales.
• Effective extraction and fusion of spatial-channel semantic features from SSVEP signals were achieved.

Conclusions:

• SED-xLSTM offers a significant advancement in SSVEP-based BCI spellers, particularly for individuals with aphasia.
• The proposed architecture effectively leverages both time and frequency domain information for improved BCI performance.
• This study highlights the potential of xLSTM and spatial attention in analyzing complex neural signals for BCI applications.