CT-Based Body Composition Measures and Systemic Disease: A Population-Level Analysis Using Artificial Intelligence Tools in Over 100,000 Patients.

Abstract

BACKGROUND. CT-based abdominal body composition measures have shown associations with important health outcomes. Advances in artificial intelligence (AI) now allow deployment of tools that measure body composition in large patient populations. OBJECTIVE. The purpose of this study was to assess associations of age, sex, and common systemic diseases with CT-based body composition measurements derived using a panel of fully automated AI tools in a population-level adult patient sample. METHODS. This retrospective study included 140,606 adult patients (67,613 men and 72,993 women; mean age, 53.1 ± 17.6 [SD] years) who underwent abdominal CT at a single academic institution between January 1, 2000, and February 28, 2021. CT examinations were not restricted on the basis of patient setting, clinical indication, or IV contrast media use. Thirteen fully automated AI body composition tools quantifying liver, spleen, and kidney volume and attenuation; vertebral trabecular attenuation; skeletal muscle area and attenuation; and abdominal fat area and attenuation were applied to each patient's first available abdominal CT examination. EHR review was performed to identify common systemic diseases, including cancer, cardiovascular disease (CVD), diabetes mellitus (DM), and cirrhosis, on the basis of relevant ICD-10 codes; 64,789 patients (46.1%) had at least one systemic disease diagnosed. Multiple linear regression models were performed for the 118,141 patients (84.0%) with no systemic disease or a single systemic disease, to assess age, sex, and the presence of systemic disease as predictors of body composition measures; effect sizes were characterized using the unstandardized regression coefficient B. RESULTS. Multiple linear regression models using age, sex, and systemic disease as predictors were overall significant for all 13 body composition measures (all p < .001) with variable goodness of fit (R² = 0.03-0.43 across models). In the models, age was predictive of all 13 body composition measures; sex, 12 measures; cancer, nine measures; CVD, 11 measures; DM, 13 measures; and cirrhosis, 12 measures (all p < .05). CONCLUSION. Age, sex, and the presence of common systemic diseases were predictors of AI-derived CT-based body composition measures. CLINICAL IMPACT. An understanding of the identified associations with common systemic diseases will be critical for establishing normative reference ranges as CT-based AI body composition tools are developed for clinical use.