EC-Transformer: Connectivity-Informed Embeddings and Adaptive Gating for fNIRS.

Abstract

Functional Near-Infrared Spectroscopy (fNIRS) provides a non-invasive modality for monitoring brain activity, yet jointly modeling temporal dynamics and inter-regional interactions remains challenging for accurate brain-computer interface (BCI) decoding. This study proposes an Effective Connectivity Transformer (EC-Transformer), which integrates connectivity-informed representations into transformer-based modeling of fNIRS signals. The architecture combines a time- wise embedding that captures temporal dynamics using positional encoding and bidirectional LSTMs with a connectivity-based embedding that encodes low-frequency directed dependency patterns. An adaptive gating mechanism dynamically fuses these representations during classification. The model was evaluated using leave-one-subject-out validation on two public fNIRS datasets involving mental arithmetic and motor execution tasks, achieving accuracies of $76.83 \pm 2.4$ and $76.03 \pm 2.00$, respectively. The proposed framework demonstrates competitive performance relative to existing transformer-based approaches while maintaining substantially lower model complexity (approximately 0.7 M parameters compared to 1.7M-3.5 M in prior models). Ablation and control analyses further suggest that EC-based embeddings provide connectivity-informed representations that complement temporal modeling while maintaining competitive decoding performance. Interpretability analyses revealed task-related connectivity patterns broadly consistent with distributed cognitive and motor-related networks. Overall, the findings suggest that incorporating connectivity-informed representations can provide physiologically structured complementary information for transformer-based fNIRS decoding while maintaining competitive performance and computational efficiency.