AI-driven appendicular skeletal muscle mass index (ASMI) prediction and low muscle mass detection from routine hip X-rays: a novel opportunistic case finding tool.
Authors
Affiliations (8)
Affiliations (8)
- School of Medicine, National Defense Medical University, Taipei, Taiwan, ROC.
- Department of Medical Education, Taichung Veterans General Hospital, Taichung, Taiwan, ROC.
- Department of Ophthalmology, Changhua Christian Hospital, Changhua, Taiwan, ROC.
- Department of Orthopedics, Wan Fang Hospital, Taipei Medical University, Taipei, Taiwan, ROC.
- Department of Orthopedics, School of Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan, ROC.
- Graduate Institute of Biomedical Materials and Tissue Engineering, College of Biomedical Engineering, Taipei Medical University, Taipei, Taiwan, ROC.
- Department of Orthopedics, Wan Fang Hospital, Taipei Medical University, Taipei, Taiwan, ROC. [email protected].
- Department of Orthopedics, School of Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan, ROC. [email protected].
Abstract
Sarcopenia diagnosis requires identifying low muscle mass (LMM), typically via dual-energy X-ray absorptiometry (DXA). However, DXA's limited accessibility restricts large-scale screening. This retrospective study aimed to develop and validate a deep learning model to predict DXA-derived ASMI from routine hip radiographs for opportunistic LMM case finding. We included a selected hospital-based cohort of 1267 patients with both hip radiography and DXA scans, split into development (n = 1140) and temporal validation (n = 127) sets. A multimodal model integrating radiographic images (ResNet-34 backbone) and clinical variables (age, sex, height, weight, and BMI) was trained to predict continuous ASMI and classify LMM per the Asian Working Group for Sarcopenia (AWGS) 2019 criteria. On temporal validation, the model achieved strong performance with Pearson r = 0.806, R<sup>2</sup> = 0.631, MAE = 0.414 kg/m<sup>2</sup>, and AUC = 0.874 for LMM classification. Applying AWGS diagnostic thresholds yielded sensitivity of 70.5% and specificity of 83.3%, with consistent performance across sex and age subgroups. Gradient-weighted Class Activation Mapping confirmed the focus on clinically relevant gluteal and proximal thigh muscles. This deep learning approach enables automated LMM identification from routine hip radiographs, offering a cost-effective, accessible tool for opportunistic LMM case finding and early intervention in at-risk populations.