Can AI improve how we handle obesity care?

A new artificial intelligence tool can predict the size of a person's waistline by simply analyzing their age, height, weight, ethnicity, and level of education, Johns Hopkins University engineers have found.

The tool's striking accuracy could help doctors estimate a patient's risk of diabetes, heart disease, stroke, and other obesity-related conditions often assessed using the infamous body mass index, or BMI, a calculation of a person's height and weight.

The findings are newly published in Diabetes & Metabolic Syndrome: Clinical Research & Reviews.

"Waist circumference is closely linked to health risks like diabetes and heart disease, but it's not regularly measured in the clinic," said corresponding author Rama Chellappa, Bloomberg Distinguished Professor of Electrical and Computer Engineering and Biomedical Engineering. "Our method makes it easier for doctors to predict a patient's obesity risk without needing to directly measure their waist, which can save time and improve the accuracy of risk assessments for obesity-related conditions."

Developed by researchers in the Artificial Intelligence for Engineering and Medicine Lab at Johns Hopkins, the highly accurate machine learning method predicts waist circumference without physical measurement. The innovative approach correctly estimates waist circumference within a narrow range about 95% of the time, offering a reliable tool for assessing obesity-related health risks.

The study was led by biomedical engineering doctoral student Carl Harris. Prasanna Santhanam, associate professor in the Division of Endocrinology, Diabetes, and Metabolism at the Johns Hopkins University School of Medicine, and Daniel Olshvang, a biomedical engineering doctoral student, also contributed to the study. The team says their work demonstrates the promise of integrating AI predictions into clinical practice, especially for treating obesity.

When assessing obesity risks, doctors often refer to a person's body mass index (BMI), a calculation of their height and weight. But BMI measurements aren't comprehensive; they don't consider body composition, ethnic differences, age, and other factors that provide a more accurate picture of a person's health. Someone with a "normal" BMI could have a higher risk of obesity-related health problems than someone with a high BMI.

Growing evidence shows that waist circumference is a better predictor of health problems related to obesity than BMI alone. However, the researchers point out that despite its predictive value, waist circumference measurement faces two challenges: the lack of a standardized measuring technique and its infrequent use in clinical practice. The Hopkins team set out to overcome those challenges.

They analyzed health data from two major studies, the National Health and Nutrition Examination Survey (NHANES), and Look AHEAD (Action for Health in Diabetes), that included patient information such as height, weight, age, ethnicity, and level of education (a surrogate for eating habits). Then they applied a machine learning technique, called "conformal prediction," to predict waist circumference. Along with the prediction, their model produces a range of values that expresses the model's confidence in the prediction's accuracy.

The team's novel approach significantly outperformed current machine learning methods to predict waist circumference. On top of that, the authors show that the uncertainty ranges were reliable and generalizable, meaning the model can make accurate predictions about populations that differ substantially from those the model was trained on, such as patients with diabetes.

The researchers emphasize that the new algorithm's ability to quantify its own uncertainty is key not only to their model's success but also for building trustworthy AI systems.

"Our approach stands out because we didn't just provide a single prediction for waist circumference—we created a range of values that show how certain or uncertain the prediction is. This adds a layer of safety and accuracy, especially in clinical settings where such uncertainty is critical and guides decision-making," said Harris.

While compelling, the researchers caution that the results are preliminary. The team said they will conduct further testing of the model across various populations and clinical settings to confirm its effectiveness in real-world situations. They plan to refine the model by including other factors like diet and physical activity, which could make predictions even more precise.