Uncertainty-Aware Multi-Outcome Screening for Undiagnosed Cardiometabolic Disease in the United States Population Using NHANES

Millions of adults in the United States live with serious health conditions they do not know they have. Diseases like diabetes, high blood pressure, and high cholesterol often develop silently, changing the body's chemistry long before a person feels sick or notices a symptom. By the time these issues are discovered, they may have already caused damage to the heart, kidneys, or blood vessels. Public health officials have long known that finding these hidden conditions early is crucial, but screening the entire population is a massive logistical challenge. Doctors cannot test everyone for every disease at every visit, and simple questionnaires often miss the mark. The core difficulty lies in balancing accuracy with trust: a computer program might predict who is at risk, but if the program is unsure of its own answer, a doctor cannot act on it without risking unnecessary panic or missed diagnoses.

A team of researchers set out to solve this problem by building a new kind of screening system that does not just guess who is sick, but also knows when it is guessing. Using data from a massive national health survey, they trained computer models to look for signs of five different undiagnosed conditions. What made their approach unique was that they added a layer of "uncertainty" to the predictions. Instead of forcing a yes-or-no answer for every person, the system was designed to admit when it was not confident enough to decide. This allowed the researchers to separate people who could be safely cleared from those who needed further testing, creating a safety net that standard computer models often lack.

The researchers turned to the National Health and Nutrition Examination Survey, a long-running study that interviews and examines thousands of Americans to track the nation's health. They focused on data collected between 2011 and 2014, assembling a group of more than 20,000 adults. Their goal was to find people who met the medical criteria for a disease but had never been told they had it by a doctor. They defined five specific targets: undiagnosed diabetes, undiagnosed high blood pressure, undiagnosed high cholesterol, a risk for kidney disease, and a combined category for anyone with at least one of these hidden issues. To ensure their computer models were truly screening for disease and not just using the test results themselves as a clue, they strictly forbade the models from using the specific blood test that defined the disease as a clue to predict it. For example, to predict undiagnosed diabetes, the model could see a person's age and weight, but it was not allowed to see their blood sugar level, which is the very thing used to diagnose the disease.

To test how well these models would work in the real world, the team simulated three different settings where a doctor might use them. The first setting was a community survey, where only basic information like age, income, and medical history was available. The second setting added routine physical exam data, such as blood pressure and waist size. The third setting, representing a full clinical visit, included common lab results like cholesterol levels. They tested six different types of computer algorithms, ranging from simple statistical formulas to complex artificial intelligence networks that learn by finding patterns in data. The results showed that as more detailed information became available, the models got significantly better at spotting the hidden diseases. When the models had access to full lab results, the most advanced tree-based algorithms could correctly identify nearly all cases of undiagnosed high cholesterol and high blood pressure, with AUC scores reaching 0.984 and 0.977 respectively.

However, the most important discovery was not just how accurate the models were, but how they handled their own confidence. The researchers used a method called conformal prediction to attach a measure of certainty to every single prediction. This system guarantees that when the model says it is confident, it is right at least 90 percent of the time. When they applied this to their results, a striking difference emerged between the models. The most sophisticated tree-based models were able to make confident decisions for almost everyone, flagging only a tiny fraction of people as uncertain. In contrast, a simpler model based on standard statistics was so unsure of its answers that it had to flag nearly 60 percent of the people for further testing just to maintain the same level of safety. This meant that while both models might look equally good on paper when measuring pure accuracy, the simpler one would overwhelm a clinic with referrals, while the smarter one could efficiently sort patients.

The study also revealed that the hidden burden of disease in the United States is substantial. When the researchers applied their findings to the national population, they estimated that about one in four American adults has at least one of these cardiometabolic conditions without knowing it. The most common hidden issue was high cholesterol, affecting roughly 13.7 percent of the population, followed closely by high blood pressure at 13.5 percent. Undiagnosed diabetes was less common but still significant at 3.1 percent. These numbers confirm that a large portion of the population is walking around with silent risk factors that could be managed if only they were detected early.

The researchers also explored how the computer models "thought" about the data. They looked at the internal representations created by their neural networks and found that these models did not see disease as a set of separate, distinct boxes. Instead, the models learned that cardiometabolic health exists on a continuous spectrum. A person might have a little bit of high blood pressure and a little bit of high cholesterol, and the model understood this as a single, flowing burden of risk rather than a collection of unrelated problems. This insight suggests that these conditions are deeply interconnected, accumulating over time in a way that simple checklists might miss.

Ultimately, the study demonstrates that the future of medical screening lies not just in building smarter algorithms, but in building systems that know when to stop and ask for help. By combining high accuracy with a clear signal of uncertainty, these tools can help doctors decide who needs a quick check-up and who needs a full workup. The research suggests that a staged approach works best: starting with simple questions in the community to find those who need closer attention, then moving to physical exams and lab tests for those flagged as high risk. This method ensures that resources are used efficiently, protecting the most vulnerable people from being missed while avoiding unnecessary testing for those who are healthy. The work provides a practical blueprint for how to use artificial intelligence in medicine not as a replacement for doctors, but as a reliable partner that knows its own limits.