Toward Trustworthy Collaborative BCI: Uncertainty-Aware Graph Reasoning for Dual-Brain Multimodal Target Detection

Rapid serial visual presentation (RSVP)-based brain-computer interfaces (BCIs) have shown strong potential for target detection in complex visual search tasks. Collaborative BCIs can further improve decision robustness by combining information from multiple users. However, most existing methods focus mainly on aggregating more evidence, while paying less attention to how the exchanged information affects other sources and whether it remains beneficial after propagation across subjects and modalities. In dual-brain multimodal RSVP, local prediction confidence does not necessarily reflect collaborative value: a locally confident source may still propagate misleading information, whereas an uncertain source may retain useful information. To address this gap, we propose DBMMNet, an uncertainty-aware dual-brain multimodal framework that moves collaborative BCI from indiscriminate evidence fusion toward trustworthy collaboration. DBMMNet combines node representations and predictive entropy to estimate sender-level propagation reliability. The reliability estimator is supervised by the counterfactual propagation utility of outgoing messages, and the resulting reliability regulates inter-source message passing. Experiments on 15 predefined two-subject groups under cross-block evaluation show that DBMMNet consistently outperforms competing methods and achieves improved robustness and stability. These findings recast trustworthy collaborative BCI from an evidence-aggregation problem into a propagation-control problem. Our code is available at: https://github.com/BillySturate/DBMMNet.