TADM: Unified Pre-trained Framework for EEG-to-fNIRS Cross-Modal Generation in BCI Decoding

Employing electroencephalography (EEG) and functional near-infrared spectroscopy(fNIRS) signals jointly is pivotal for advancing brain-computer interface (BCI) decoding, yet physical clashes between EEG and fNIRS channels prevent true simultaneous capture. Since EEG is cheap and easy to collect, we can sidestep the hardware conflict by deriving the missing fNIRS traces from the recorded EEG. The core challenge is to learn this mapping in a way that transfers across subjects, tasks and recording rigs without retraining. To address this issue, this work proposes the Pretrained Unified Framework: Topology-Adaptive Diffusion Model(TADM) for EEG to fNIRS Cross-Modal Generation in BCI-decoding. Trained on larger and more diverse datasets, TADM aims to enhance feature extraction capabilities and construct a more generalizable modality mapping mechanism. To address the heterogeneity between EEG and fNIRS signals, the model adopts a unified data alignment strategy for effective alignment and incorporates a pretrained model that leverages the geometric configuration of EEG and fNIRS sensors to construct a unified representation. The final cross-modal generation is achieved via a conditional diffusion process that transforms EEG latent representations into corresponding fNIRS latent representations. Experimental results demonstrate that TADM exhibits strong robustness and broad adaptability across diverse task scenarios and sensor configurations.