An AI-Based Prediction of Cognitive Capacity in Older Adults and Individuals With Mild Cognitive Impairment During Virtual Reality Driving Tasks

Driving ability in older adults is closely associated with independent mobility and social participation. However, under conditions of high cognitive load or distraction, visual attention failures-such as the "look-but-fail-to-see" phenomenon-frequently occur and substantially increase crash risk. Previous studies have demonstrated that older adults and individuals with mild cognitive impairment (MCI) exhibit declines in visual attention allocation, executive control, and visuomotor transformation efficiency. These dynamic regulatory processes are difficult to capture using conventional paper-and-pencil or static neuropsychological assessments, underscoring the need for ecologically valid and dynamic evaluation approaches.

Virtual reality (VR), when integrated with eye-tracking and upper-limb motion analysis, enables the simulation of realistic driving environments under safe and controlled conditions. This approach facilitates precise quantification of visual search behavior, hazard detection, and visuomotor coupling, thereby offering a novel framework for cognitive assessment and driving risk prediction. Despite these advances, three critical gaps remain in the current literature: (1) a lack of systematic investigations focusing on older adults and individuals with MCI across diverse VR driving scenarios to examine visual search and attentional control; (2) artificial intelligence (AI) models that are predominantly limited to single tasks or single modalities, restricting their ability to generalize across contexts; and (3) insufficient integration of upper-limb operational data to fully characterize the dynamic interactions among vision, action, and cognition.

To address these gaps, this study adopts a multi-phase, multi-level research design. In Year 1, a VR-based eye-movement system incorporating controllable cognitive load will be developed to examine pro-saccade and anti-saccade performance among young adults, cognitively healthy older adults, and individuals with MCI. This phase will establish a high-sensitivity, eye-movement-based AI model for MCI identification. In Year 2, the framework will be extended to multi-scenario VR driving tasks through the synchronous integration of eye-tracking and upper-limb operational data, enabling characterization of visuomotor coupling under varying cognitive demands. External validation, transfer learning, and multimodal fusion techniques will be applied to enhance cross-scenario generalizability. In Year 3, multimodal datasets will be integrated to develop an explainable artificial intelligence (XAI) prediction system. Longitudinal follow-up will be conducted to evaluate its prognostic validity for changes in cognitive and driving performance, ultimately leading to a clinically applicable decision-support prototype.

At the theoretical level, this study aims to elucidate the mechanisms underlying visual attention, working memory, executive control, and visuomotor coupling in older adults and individuals with MCI under dynamic conditions. At the clinical and practical levels, it seeks to develop a non-invasive and repeatable digital cognitive screening tool for early MCI detection and older-driver risk assessment, as well as to provide evidence-based support for traffic safety policy development.