Interpretable digital multimodal evaluations for sarcopenia in aging populations

Background The multimodal nature of sarcopenia complicates its assessment in older adults. The Asian Working Group for Sarcopenia (AWGS) 2025 consensus simplifies diagnosis by requiring concurrent low muscle mass and strength, improving clinical utility. However, quantitative and interpretable assessments remain constrained by limited well-normalized data and challenges in multimodal alignment. Methods Here, we introduce three-level balancing assessment (TLBass), an interpretable AI model integrating multimodal sensing data with human dynamics models to evaluate sarcopenia. Data from 355 older adults in Hong Kong were collected for model development and validation. Results TLBass achieved an area under the curve (AUC) of 0.92 for surrogate evaluations in the internal dataset. Only its behavioral base model (SBA) underwent external validation, independently achieving an AUC of 0.81 in an external clinical dataset, supporting its potential for rapid clinical screening. Conclusions TLBass provided interpretable associations between sensing data and clinical outcomes, including surrogate markers of fast-twitch muscle fiber characteristics, declining muscle strength, and impaired balance. These findings highlight the potential of interpretable AI to support personalized sarcopenia assessment.