Predicting Body Composition from Chest Radiographs by Deep Learning: 10-year Mortality and Geriatric Outcomes

Abstract

BackgroundBody composition strongly influences clinical outcomes in older adults, yet body mass index (BMI) lacks discriminatory power, and standard tools such as bioelectrical impedance analysis (BIA), dual-energy X-ray absorptiometry are not routinely accessible. Deep learning enables scalable, opportunistic assessment of body composition from chest radiographs (CXRs), one of the most widely available imaging modalities.

Methods and FindingsUsing the Inception-V3 architecture, we developed a deep-learning model using 107,568 paired CXR and BIA records (2016-2018). The model was temporally validated on a separate dataset of 77,655 records (2014-2015). Our model predicted skeletal muscle mass (SMM) and fat mass (FM) with high accuracy (SMM: Pearson r = 0.967, MAE 1.40 kg; FM: r = 0.924, MAE 1.61 kg). In a cohort of 5,932 older adults (aged [≥]65years), a 1-SD increase in CXR-predicted skeletal muscle index (SMI) was associated with a significant reduction in 10-year all-cause mortality (Hazard Ratio [HR] 0.65 [95% CI 0.58-0.73] for men; 0.80 [0.67-0.97] for women). In an external validation of 925 geriatric clinic patients, predicted SMI also showed comparable associations with geriatric parameters, including lower odds of sarcopenia (per 1 SD increase: 0.29 [0.22-0.38] for men; 0.25 [0.18-0.34] for women) and frailty (0.62 [0.48-0.78] for men; 1.00 [0.81-1.23] for women). These associations were more robust than those of BMI. Key limitations include the retrospective, single-center design and the use of a relatively healthy screening population.

ConclusionA deep learning model applied to routine CXRs enables accurate estimation of skeletal muscle and fat mass, demonstrating prognostic and functional relevance comparable to BIA measurements. This approach may serve as a practical, low-cost tool for risk stratification and long-term care planning, particularly in older adults.