AI-based multimodal behavioral phenotyping for sarcopenia and muscle health assessment in older adults: study protocol

Abstract Background Sarcopenia, characterized by age-related declines in muscle mass and strength, is associated with functional decline in older adults. Objective assessment may be impractical for frequent community monitoring, while brief self-report tools such as the SARC-F have limited sensitivity. Facial dynamics, vocal characteristics, and movement may carry information outside conscious control, yet few studies have tested whether such phenotypes distinguish older adults with normal muscle mass and strength from those with low muscle mass. This protocol therefore frames community assessment in three linked layers: standard-instrument responses, a mediating layer accommodating varying digital familiarity, and multimodal behavioral signals generated throughout. The mediating layer is currently performed by a trained facilitator whose interaction patterns can inform the design of future embodied conversational agents (ECAs). Methods This study is conducted within an ongoing interdisciplinary sarcopenia research program at Yeungnam University. Up to 120 community-dwelling adults aged ≥ 65 years are recruited independently of the parent program. Each participant completes a single session of under 30 minutes. The session is mediated throughout by a trained facilitator, and facilitator–participant interaction patterns are systematically documented for descriptive analysis. The session comprises warmup conversation, a read-aloud passage, facial expression tasks, open-ended emotion elicitation, STAI-6, SGDS-K, SARC-F, and selected SarQoL-K, with continuous audiovisual recording. Verbal and nonverbal features, grounded in behavioral and psychosocial theory, are analyzed using interpretable machine-learning models and exploratory multimodal fusion to examine their potential for human-centered AI assessment of muscle health. For group comparison, appendicular skeletal muscle mass index (ASMI) by bioelectrical impedance analysis and handgrip strength (HGS) by dynamometry are used to classify participants according to AWGS 2019 thresholds. Discussion This protocol serves a dual purpose. First, it evaluates whether multimodal behavioral phenotypes are comparable to objective muscle-strength and body-composition measures. It also examines whether these phenotypes provide information complementary to self-report screening and may inform the development of non-invasive approaches to community muscle-health assessment. Second, the facilitator-mediated design documents interaction patterns that may inform future ECAs for older adults. The framework provides a structured approach to multimodal data collection across varying levels of digital familiarity. Trial registration: Not applicable (observational study).