Multimodal AI Model for Sarcopenia Detection: Integrating Chest CT and Clinical Parameters in Older Adults

Sarcopenia, an age-related syndrome marked by progressive loss of skeletal muscle mass and function, is associated with frailty, disability, falls, and increased mortality among older adults. However, existing diagnostic methods, such as dual-energy X-ray absorptiometry (DXA) and physical performance tests, are often inaccessible in routine clinical practice due to equipment and time constraints.

This study aimed to develop and validated a multimodal, explainable AI model for identifying sarcopenia using routinely available chest CT scans and outpatient clinical data in older adults.

A total of 290 participants (mean age 67.6 ± 5.8 years; 38.9% female) were included. A weakly supervised segmentation framework combining the Segment Anything Model (SAM) and Contrastive Language-Image Pretraining (CLIP) was employed to extract muscle features at the T12 level. Clinical variables, including anthropometric indices, lifestyle behaviors, and biochemical markers, were encoded and fused with imaging-derived features. A multi-layer perceptron (MLP) was trained to classify sarcopenia based on 2019 AWGS criteria. Model interpretability was assessed using SHAP (Shapley Additive Explanations) values.

The model achieved an AUC of 0.88 (95% CI: 0.83-0.92), accuracy of 0.85 (95% CI 0.82-0.89), sensitivity of 0.79 (95% CI: 0.70-0.987), and specificity of 0.88 (95% CI: 0.83-0.92). SHAP analysis revealed that gender, total triiodothyronine, creatine kinase, body mass index and creatinine were the most influential predictors. The fusion of weakly supervised learning and multimodal data enabled effective muscle region segmentation and improved diagnostic performance.

In summary, we developed and internally validated an explainable multimodal AI model that integrates chest CT-derived muscle features with routine outpatient clinical variables for sarcopenia detection in older adults. The model demonstrated strong diagnostic performance and interpretability, highlighting its potential for opportunistic risk stratification in routine clinical workflows. Future multi-center validation and prospective studies are warranted to confirm its generalizability and long-term clinical utility.