August 18, 2026 | The worried well the world over now have readily accessible and easy to self-administer home tests for assessing their risk for diabetes and chronic kidney disease, thanks to the computational skills of a pair of researchers at DTU Sustain, based at the Technical University of Denmark. The new online tools focus on a small number of parameters identified by artificial intelligence (AI) as being the most important among hundreds of possibilities in the long-running National Health and Nutrition Examination Survey (NHANES) in the U.S., according to postdoc researcher Daniel Yoo, Ph.D. who has been leading the charge.
Among the risk indicators are the usual suspects, including age and body mass index (BMI). Intriguingly, the list also includes a few surprises—notably, thigh length and arm circumference for diabetes and upper arm length for chronic kidney disease. For the latter, an individual’s poverty-income ratio also turns up as a factor depending on which clinical equation is used to define kidney health, highlighting how socioeconomic status is intertwined with chronic disease risk, says Yoo.
The first of a trio of studies, introducing the Machineborne Early Diabetic Warning and Control System (MEDWACS), was published recently in the Journal of Clinical Epidemiology (DOI: 10.1016/j.jclinepi.2026.112266). This was immediately followed by a research article about the Machineborne Early Renal Warning and Control System (MERWACS), which appeared in PLOS Digital Health (DOI: 10.1371/journal.pdig.0001486). A third study, currently under review, investigates whether integrating detailed dietary history—specifically, adherence to the Healthy Eating Index—improves the MEDWACS diabetes screening model.
Umberto Maggiore, M.D., kidney specialist at the University of Parma, serves as the clinical expert on the research team. Yoo and his collaborating supervisor, Olivier Jolliet, Ph.D., admit they were particularly skeptical of the thigh length parameter (longer being better), but Maggiore found the measurement biologically interesting with two plausible explanations.
The first is that it might be related to malnutrition in early childhood, stunting a person’s growth, relays Yoo. The other, more fascinating possibility is that thigh length is a “proxy measure to represent muscle mass.”
The thigh contains the largest overall muscle group in the body, and muscles consume blood glucose, he explains. Greater muscle mass improves insulin sensitivity and helps naturally buffer blood sugar levels.
This was but one of seven parameters in MEDWACS, the others being systolic blood pressure, upper arm circumference (smaller being better), age, gender, BMI, and waist circumference. MERWACS, intended for adults aged 50 and older, comprises a slightly longer list of 12 non-invasive parameters, inclusive of history of heart failure, history of diabetes, and history of hypertension as well as ethnicity, upper arm length, diastolic blood pressure, and 60-second pulse.
Both risk prediction tools were built using 30 years of survey data, much of it not the kind of information typically found in clinical datasets. Other than standing height and weight, they contain body measures that aren’t routinely collected by hospitals and clinics, says Yoo. “That’s a pitfall for data scientists developing a machine learning model who want to include as many variables as possible ... to increase the predictability.”
Fortunately, Yoo had practice focusing on a smaller number of parameters while working on his Ph.D. in Paris where he developed AI-driven solutions and a virtual biopsy system in kidney transplantation that needed to be easy for physicians and clinicians to use. “I converted this concept to a more general public,” he says. “They can predict their own risk ... to have an earlier warning,” and thereby reduce the cost burden associated with caring for disease occurrences at later, more severe stages.
From an initial set of nearly 3,700 possible options in the NHANES dataset, Yoo and his team used AI to whittle down the list to roughly 100 of the most important parameters for the MEDWACS tool and then started filtering out the noise. The process began by eliminating duplicate responses as well as new parameters introduced or whose definitions had changed with time, says Yoo.
“We had to choose selectively only those consistently appearing for those 30 [survey] years,” he explains. While this meant possibly missing some important parameters, it also ensured having three decades of usable data.
Since some variables had a lot of missing data, those, too, were removed, continues Yoo. Including those parameters would require using imputation methods that he likens to filling in absent values with “synthetic” data that could be wrong.
A machine learning method called Boruta was used to identify the best of the remaining variables based on their predictive contribution. The final seven parameters emerged after removing many diet-related inputs where food type and consumption habits couldn’t be adequately captured, as well as blood data, primarily because the explicit goal was to create a 100% non-invasive tool that requires zero laboratory tests, Yoo says.
All seven are easy enough for someone to input at home, although there is yet no single public-facing, official government or public health domain hosting the live interactive calculator. But it can be accessed via a website maintained by the university, which includes a manual with photos to provide guidance on how to do the measurements based on guidance issued by the U.S. Centers for Disease Control and Prevention, which operates NHANES.
The entire model-producing exercise was completed in an impressive five months, since Yoo was repurposing the process and working with a publicly available dataset. “I just had to manage the dataset and understand it better,” he says.
MEDWACS has been validated in two external validation tests as well as internally using different portions of the NHANES dataset for training and testing the AI model, reports Yoo. In real-world validation studies using NHANES data from a different timeline than the test set, as well as a South Korea dataset, the model in both instances demonstrated strong performance relative to other recognized testing methods. “Because it performed consistently across two vastly different populations, we are confident it will work well in many different settings. However, as with any medical AI, we would want to validate it on local data before deploying it in completely new regions.”
Screening guidelines issued by the U.S. Preventive Services Task Force and the American Diabetes Association related to age (over 35) and BMI (25 or higher) are the standards against which MEDWACS was compared in the published study, he says. But other diabetes self-screening tools exist, notably the seven-question, point-based Leicester Risk Assessment score, endorsed by Diabetes UK (age, gender, ethnicity, BMI, waist circumference, and family history of diabetes and hypertension).
There is both a clinical need for tests like these as well as worrying behavioral signs in the population, says Yoo. “We know that diabetes is increasing because of our [frequently sedentary] lifestyle and ... especially in the United States, the obesogenic environment” marked by an abundance of fast-food chains.
The added problem in the U.S. is that healthcare is not free, so for economic reasons people may be reluctant to visit the doctor, he says. But even in Europe, which has comparatively fewer cost concerns, many people aren’t worried enough about their health to take well-established preventive steps related to diet, exercise, and getting preventive screening tests.
At the prediabetes stage, the condition is often controlled via weight loss, more physical activity, and a balanced diet, continues Yoo. But it is much more difficult to reverse once a person transitions to diabetes. His top advice for people finding themselves in the danger zone is to modify the parameters they have control over—notably, weight, waist circumference, and blood pressure—which are major predictors of their future health.
Development of the MEDWACS and MERWACS tools have been “purely academic” undertakings, says Yoo. The work has been funded by the Novo Nordisk Foundation, the independently operated parent entity of the Danish multinational pharmaceutical company that manufacturers medications for both diabetes and chronic kidney disease.