From egocentric data to candidate cognitive markers: AI modelling with Meta Aria glasses in daily-living Scenarios

Le résumé fourni par la source

Cognitive screening is typically episodic and confined to clinical settings; however, early functional changes may first emerge during everyday activities and remain undetected in formal assessment. To examine such changes under realistic conditions, this study adopts a Living Lab approach, collecting Meta Aria first-person (egocentric) recordings during controlled daily-living tasks. Participant feedback and recording review informed successive protocol refinements and data- quality safeguards. The approach was evaluated in 30 clinician-labelled older adults (20 controls; 10 with mild or subjective cognitive impairment [MCI/SCI]) through two exploratory modelling paths governed by a shared evaluation contract. Path A uses learned multimodal representations to predict a 7-test cognitive composite (R² = 0.417) and distinguishes the MCI/SCI group from controls at balanced accuracy 0.80 [0.63–0.95], above an age + education + sex baseline (0.68) and below a neuropsychological-test comparator (0.85). Path B derives interpretable behavioural markers from head-motion, trajectory and gaze features, producing an auditable inventory of candidate cognitive markers. Together, the paths suggest that egocentric daily-living data contain cognition-relevant signal, that Living Labs can de-risk emerging egocentric AI for cognitive health, and that interpretable markers may support downstream Ambient Assisted Living applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.