BELLEVUE, Wash., Oct. 29, 2025 (GLOBE NEWSWIRE) — Truveta is proud to announce the publication of its latest peer-reviewed research in Radiology Advances: “XComposition: Multimodal Deep Learning Model to Measure Body Composition Using Chest Radiographs and Clinical Data.” This groundbreaking study demonstrates the power of artificial intelligence to estimate critical body composition measures—such as visceral and subcutaneous fat volumes—from a simple chest radiograph combined with commonly available clinical data. The deep learning model is available as a Python library for others to experiment with in Truveta’s GitHub.
Key findings
The research team developed a multimodal deep learning model that integrates chest radiographs (CXR) with four basic clinical variables (age, sex at birth, height, and weight) to estimate body composition typically measured by CT scans. The study analyzed data from more than 1,100 patients across a subset of Truveta member health systems in the US.
- The multimodal model accurately estimated subcutaneous fat volume (Pearson’s R: 0.85) and visceral fat volume (Pearson’s R: 0.76).
- A late fusion strategy—combining imaging and clinical data at the decision level—yielded the best results (p < 0.04 for subcutaneous fat volume).
- The multimodal model outperformed imaging-only and clinical-only approaches across all key body composition metrics (p < 0.001 for subcutaneous fat volume).
Why it matters
Body composition is an important predictor of cardiovascular disease, diabetes, and cancer prognosis. Traditional methods to measure these metrics—such as MRI or CT—are expensive, resource-intensive, and not always accessible to patients. This study shows that a chest radiograph, one of the most common and widely available imaging tests, can serve as a low-cost, scalable tool for estimating body composition when combined with AI.
“Our work shows that we can unlock clinically meaningful insights from a chest X-ray—an exam that millions of people receive each year,” said Ehsan Alipour, MD, PhD, a machine learning post-doctoral researcher at Truveta and lead author of the study. “By combining imaging with just a few simple clinical variables, we created a powerful, accessible way to estimate body composition that could help improve screening, research, and ultimately patient care.”
About the study
This study leveraged Truveta Data, the most complete, timely, and representative dataset of de-identified electronic health records (EHR) in the US, contributed by a collective of leading health systems. Imaging data were linked with clinical variables across health systems, enabling the development and validation of this multimodal AI model.
The full paper is available in Radiology Advances.
About Truveta
Truveta is a real-world intelligence company transforming medical science with unprecedented data and predictive AI. We power breakthrough discoveries, accelerate regulatory-grade evidence, and unlock real-time insights from a dataset uniquely built with and owned by US health systems—united by a mission of Saving Lives with Data.
Truveta membership includes Providence, Advocate Health, Trinity Health, Tenet Healthcare, Northwell Health, AdventHealth, Baptist Health of Northeast Florida, Baylor Scott & White Health, Bon Secours Mercy Health, CommonSpirit Health, Hawaii Pacific Health, HealthPartners, Henry Ford Health System, HonorHealth, Inova, Lehigh Valley Health Network, MedStar Health, Memorial Hermann Health System, MetroHealth, Novant Health, Ochsner Health, Premier Health, Saint Luke’s Health System, Sanford Health, Sentara Healthcare, Texas Health Resources, TriHealth, UnityPoint Health, Virtua Health, and WellSpan Health.
- Truveta scientists developed and validated a multimodal deep learning model using chest radiographs and de-identified electronic health records from Truveta Data.
- Multimodal AI outperformed single-source models. The late fusion multimodal model achieved the best performance across nearly all body composition measures, significantly outperforming imaging-only and clinical-only approaches.