VR rehabilitation system evaluator: A fNIRS-based and LLM-enabled evaluation paradigm for Mild Cognitive Impairment

Abstract

Virtual Reality (VR) has significant potential for Mild Cognitive Impairment (MCI) rehabilitation by offering dynamic and engaging tasks designed to stimulate various cognitive domains. However, how efficiently evaluating the VR tasks is always challenging. The current effectiveness evaluation methods for task evaluation and optimisation typically rely on post-training metrics like game performance or subjective questionnaires, which cannot directly reflect the comprehensiveness and intensity of cognitive stimulation. To bridge these gaps, we propose a novel approach integrating functional near-infrared spectroscopy (fNIRS) data and a Large Language Model (LLM) for VR rehabilitation task evaluation and optimisation. Firstly, we leverage fundamental cognitive theories to propose a systematic evaluation paradigm for assessing the scope of the cognitive domains stimulated by VR tasks. Secondly, we propose fNIRS-derived graph parameters to objectively evaluate the cognitive stimulation with high-time resolution. Finally, we develop a prompt-generating strategy to facilitate Large Language Model (LLM)-based analysis, which transforms fNIRS-derived metrics combined with the scope of stimulated cognitive domains into easy-to-understand evaluation reports and actionable recommendations for task optimisation. We demonstrate this approach with two VR rehabilitation tasks. Several experts evaluated the LLM-based task evaluation results and task optimisation recommendation and agreed that the LLM-generated results are meaningful and applicable. This approach offers a rapid, objective method to evaluate VR task effectiveness and improves their design, marking a significant step in developing more effective MCI interventions.