Predicting future heart disease from hemodynamic measurements during a cold-pressor test

What is it about?

Heart disease and stroke are among the leading causes of death worldwide. In Norway, doctors use a tool called NORRISK2 to estimate a person's 10-year risk of having a heart attack or stroke, based on factors like age, blood pressure, cholesterol, and smoking. This study asked whether measuring how a person's heart rate, blood pressure, etc react to a brief, standardized stressor could make that risk prediction even better. Nearly 6,700 participants in the Tromsø Study, a long-running health survey in northern Norway, took part in a "cold-pressor test," where they placed a hand in ice-cold water for about a minute while a sensor continuously tracked their heart rate and blood pressure. Researchers then tracked these participants for several years to see who went on to develop cardiovascular disease. The team tried two approaches. First, they added summary measures of heart rhythm variation and blood pressure reflexes to the standard NORRISK2 calculator. Second, they used artificial intelligence (machine learning) to analyze the raw, second-by-second pattern of heart rate and blood pressure changes during the test, without using any other health information about the person. Adding the extra heart measurements to the standard calculator did not improve its accuracy. However, the AI approach, using only the heart rate and blood pressure patterns from the cold-water test, was still able to meaningfully predict who would go on to develop heart disease, even without knowing anything else about the person. Its predictions largely overlapped with what the standard risk calculator already captures, suggesting the two methods are picking up on similar underlying risk information. Overall, this study shows that brief measurements of how the heart responds to a physical stressor do carry real, measurable information about future heart health, information that future, larger studies might be able to put to practical use.

Featured Image

Photo by Julia Taubitz on Unsplash

Photo by Julia Taubitz on Unsplash

Why is it important?

Wearable devices that track heart rate and, increasingly, blood pressure are now used by hundreds of millions of people worldwide. This raises an important and timely question: can the everyday data these devices collect actually help predict serious health outcomes, such as heart attacks and strokes, rather than just track fitness? This study is one of the few to directly and rigorously test that question against a real, government-endorsed clinical risk calculator (NORRISK2), rather than comparing new methods only to each other in isolation. Using data from nearly 6,700 participants in the long-running Tromsø Study, followed for years afterward for actual cardiovascular events, the researchers combined a well-established statistical model with a modern machine learning approach applied to raw, beat-to-beat heart and blood pressure signals, the same kind of continuous physiological data a smartwatch or wearable sensor could one day capture. Unlike many AI health studies that report striking accuracy but lack a fair, real-world benchmark, this work transparently shows where machine learning does and doesn't add value: it can independently detect cardiovascular risk from physiological patterns alone, but it doesn't yet outperform or meaningfully improve upon existing clinical tools. That kind of honest, head-to-head result is valuable and increasingly rare, and it helps set realistic expectations for where wearable-based AI risk prediction stands today and what's needed, larger studies, better sensors, and more refined models, before it can move from research into everyday preventive care.

Perspectives