Can AI serve as a bridge between laboratory and real-world pupillometry? A perspective on real-world neuroscience

Measuring human brain activity in the real world has long remained a major challenge in conventional neuroscience. Pupil dynamics offer a promising non-invasive proxy for arousal-related multidimensional brain states. However, interpreting pupil data collected outside the laboratory is notoriously difficult due to complex, multilayered factors, including lighting conditions, arousal states, and higher-order cognitive control, that can confound results. In this perspective, I propose an AI-driven analytical strategy as a potential solution to this interpretational bottleneck. I discuss a potential approach that employs deep learning to predict and model human pupil dynamics from multimodal contexts, such as first-person vision (FPV) video and physiological and physical data as well as the application of translational data from non-human studies. This AI strategy may enable the disentanglement and extraction of internal-state factors from complex real-world pupil data, as well as the virtual estimation or prediction of internal states even in the absence of pupil data. Additionally, I discuss how this approach contributes to the NeuroAI and how it could open the door to real-world neuroscience in fields such as sports science, clinical medicine, and next-generation human-computer interaction, alongside the challenges to realizing this vision.