Online Inference of Human Intention as a Latent Control State from Single-Trial EEG

Abstract

Human intention can be modeled as a latent internal state that modulates how sensory information is evaluated and translated into action in human-machine systems. However, most existing brain-computer interfaces (BCIs) rely on control signals tightly coupled to externally imposed stimulation and do not explicitly infer whether perceived stimuli align with a user's internal goals. Here, we investigate whether intention can be inferred as a latent, goal-dependent state from single-trial electroencephalography (EEG). We introduce a stimulus-based paradigm in which intention is specified by an internally cued target category, while object identity varies independently across stimuli. To estimate intention under single-trial neural variability, we propose an interpretable fuzzy prototype-based network that maps each trial onto interpretable fuzzy prototypes encoding intention-specific dynamics. The model represents intention-related neural activity using a compact set of fuzzy prototypes with soft memberships, enabling robust decoding without reliance on engineered mediating stimuli. Experimental results demonstrate reliable within-subject single-trial intention decoding that outperforms representative deep learning baselines, achieving an accuracy of 93.22% +/- 3.21%. Online validation further confirms real-time feasibility, with an accuracy of 70.11% +/- 10.87%. Together, these findings advance intention-aware BCIs from stimulus-driven detection toward principled inference of goal-dependent internal states.