JMIR Publications

ABSTRACT

Background:

Sarcopenia affects 10%–27% of older adults and is associated with increased morbidity and functional declines. Existing screening tools rely on static assessments and fail to capture continuous daily functional behaviors. In this study, we developed an integrated multimodal framework combining wearable-derived physical activity data with clinical features for sarcopenia screening.

Objective:

This study aimed to develop and evaluate a multimodal deep learning framework integrating wearable-derived physical activity data and clinical features to enable scalable and real-world sarcopenia screening.

Methods:

One hundred community-dwelling adults aged ≥65 years (50 with sarcopenia, diagnosed according to the AWGS 2019 criteria, and 50 controls without sarcopenia) were recruited in northern Taiwan. The Sarcopenia Dynamic-Static Fusion (SarcFuse) framework integrated 7-day accelerometer-derived physical activity signals with cross-sectional clinical features (demographics, anthropometric measures, body composition, and handgrip strength) using a query-based cross-attention multimodal fusion architecture. Data were split 70/30 for training/testing with five-fold cross-validation. Performance was evaluated using the accuracy, sensitivity, specificity, F1-score, and area under the receiver operating characteristic curve (AUROC), with 95% confidence intervals (CIs).

Results:

This study included 50 participants with sarcopenia and 50 normal controls, with mean ages of 72.94 ± 4.52 years and 70.74 ± 3.38 years, respectively. Compared with controls, the sarcopenia group showed lower BMI (21.48 vs 23.17 kg/m²), lower upper-limb muscle percentage (10.60% vs 11.30%), higher upper-limb fat percentage (15.18% vs 14.39%), higher trunk muscle percentage (53.71% vs 52.76%), lower trunk fat percentage (50.44% vs 51.92%), and reduced appendicular skeletal muscle mass (ASM) (12.78 vs 15.60 kg) and skeletal muscle mass index (SMI) (5.26 vs 6.22 kg/m²) (all P<.01). For baseline comparison, maximal handgrip strength and SARC-F combined with calf circumference questionnaire (SARC-CalF) achieved AUROCs of 88.89% and 68.06%, respectively. The proposed SarcFuse framework achieved the best overall performance, with an AUROC of 96.06% (95% CI: 94.18%–97.67%), accuracy of 93.20% (95% CI: 91.20%–94.94%), F1-score of 91.31% (95% CI: 88.69%–93.68%), and recall of 89.33% (95% CI: 85.71%–92.72%), with the lower bounds of key performance metrics exceeding those of baseline screening tools. Interpretability analyses further identified clinically meaningful composite indices, including the AWGS 2019 risk index, overall muscle mass index, and fat-muscle balance, as key contributors to model predictions.

Conclusions:

This study demonstrates the feasibility of a multimodal screening framework for sarcopenia that integrates clinical and wearable-derived data. The proposed framework may support scalable risk stratification and early identification in community and clinical settings, rather than serving as a stand-alone diagnostic tool. Further multicenter external validation and prospective evaluation are required to confirm generalizability and clinical utility. Clinical Trial: This study was approved by the Institutional Review Board of Taipei Medical University (No. N202204103).