Classification of EEG-fNIRS bimodal brain signals for motor imagery tasks based on wavelet transform and...

The fusion of Electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) provides richer neural information for brain-computer interface decoding. However, due to their distinct physiological mechanisms and heterogeneous temporal and statistical properties, EEG and fNIRS are difficult to temporally align and to project into a shared latent representation. To address this challenge, we propose BiCAT, a lightweight bimodal decoding framework that integrates wavelet-based preprocessing, artifact-aware time-domain refinement, and early feature-level fusion with a compact Transformer encoder. Wavelet transform is first applied to separate signal and noise components across frequency bands, after which spatio-temporal domain processing suppresses motion and physiological artifacts while preserving task-relevant patterns. The cleaned EEG and fNIRS features are concatenated and fed into a single-encoder Transformer, where joint self-attention captures salient temporal cues within each segment.BiCAT is evaluated on two publicly available EEG-fNIRS datasets covering motor imagery (MI), mental arithmetic (MA), and word generation (WG) tasks. The model achieves 93.41 % accuracy on MI, outperforming the strongest unimodal baseline (fNIRS) by 4.39 percentage points. On MA and WG, BiCAT attains 96.47 % and 96.41 % accuracy, corresponding to gains of 10.39 and 3.86 points over the best unimodal fNIRS and HbR baselines, respectively. Despite having only 111 k parameters, BiCAT performs competitively with representative multimodal fusion methods on the same benchmarks. These results demonstrate that BiCAT provides effective bimodal feature integration and robust performance across multiple EEG-fNIRS tasks while maintaining low computational complexity.