AI Model Analyses Body Composition to Predict Health Risks 

OAK BROOK, Ill. – Researchers used AI to analyze whole-body MRI scans from more than 66,000 participants to create the most detailed reference map to date of how fat and muscle are distributed in the human body across age, sex and height. The study was published today in Radiology, a journal of the Radiological Society of North America (RSNA). Results of the study show that the quality and amount of skeletal muscle, not just visceral fat, are strong predictors of diabetes, major cardiovascular events and mortality.

Clinicians have long relied on body mass index (BMI) and body weight to estimate cardiometabolic—the connection between cardiovascular (heart/blood vessel) and metabolic (energy/nutrient processing) systems on health—and overall health risk. But BMI is a crude measure of body composition that only relies on height and weight and does not account for muscle mass or fat distribution.

“Many risk scores and treatment decisions still rely on BMI or waist circumference because they are simple to obtain,” said senior author Jakob Weiss, M.D., Ph.D., radiologist in the Department of Diagnostic and Interventional Radiology at University Medical Center Freiburg in Germany. “But BMI does not reliably reflect a person’s actual body composition.”

Dr. Weiss said the medical community also lacks reference standards for how body composition changes in asymptomatic individuals as they age, as well as differences between men and women. 

“There is growing evidence that body composition measures are independent risk factors for cardiometabolic and oncological diseases and mortality,” said first author Matthias Jung, M.D., from the Department of Diagnostic and Interventional Radiology, University Medical Center Freiburg. “However, these measures are influenced by height and sex and change substantially with age.”

The retrospective study included a cohort of 66,608 individuals (mean age 57.7 years, 34,443 males, mean BMI: 26.2) who underwent whole-body MRI as participants in the UK Biobank and the German National Cohort between April 2014 and May 2022.

The researchers calculated age-, sex-, and height-normalized body composition metrics from the MRI scans using their open-source, fully automated deep learning framework. The body composition metrics, including subcutaneous adipose tissue, visceral adipose tissue, skeletal muscle, skeletal muscle fat fraction, and intramuscular adipose tissue, were expressed as z‑scores, which show how far an individual deviated from the age-, sex-, and height-adjusted norm. 

The researchers then conducted statistical analyses to assess the prognostic value of z-score categories (low: z<-1; middle: z=-1 to 1; high: z>1) to predict the incidence of diabetes, major adverse cardiovascular events and all-cause mortality. 

They found that high visceral fat was associated with a 2.26-fold increased risk of future diabetes, high intramuscular fat was associated with a 1.54-fold increased risk of future major cardiovascular events, and low skeletal muscle was associated with a 1.44-fold higher all-cause mortality beyond cardiometabolic risk factors.

“It’s not only how much muscle you have, but also it’s the quality of that muscle,” Dr. Jung said. “Knowing the volume of intramuscular fat gives us a window into muscle quality that other methods like BMI, bioelectrical impedance analysis, or DEXA can’t easily provide.”

The research team also generated age-, sex-, and height-normalized reference curves for key body composition measures. 

“Adjusting for confounding factors is critical for improving screening accuracy and tailoring treatment decisions,” Dr. Weiss said. “This tool has the potential to identify whether an individual’s body composition puts them at greater risk for metabolic disease compared to their age-matched peers.”

The researchers released their open-source web-based age-, sex-, and height-adjusted body composition z-score calculator to support future research, accelerate clinical translation, enabling researchers and clinicians to normalize their own datasets for improved comparability and generalizability.

“This tool can allow clinicians to use routine imaging opportunistically,” Dr. Weiss said. “A dedicated whole-body MRI is not necessarily required. If a routine CT or MRI body scan already exists, the information can be extracted for benchmarking against the reference values.”

Dr. Weiss said the AI tool could also help improve risk stratification in oncology or distinguish desirable fat loss from unwanted muscle loss in patients using weight-loss drugs such as GLP‑1 agonists.

“We’re already imaging patients every day,” Dr. Weiss noted. “On every scan of the abdomen or chest, the information is there, we just don’t routinely measure or report it. AI now allows us to tap into this hidden layer of data in a quantitative, reproducible way.”

Next steps for the researchers include validating the reference curves in clinical populations, especially predicting treatment toxicity, survival and recurrence in cancer patients, and developing disease-specific reference values for other patient groups.