New Artificial Intelligence Tool Diagnoses Masked Hypertension

About 10% of people with hypertension have normal blood pressure readings at the doctor's office. Now researchers at the University of Arkansas have developed an artificial-intelligence diagnostic tool to detect this condition, known as masked hypertension.  

The work, led by associate professor of chemical engineering William J. Richardson, could help more people get treatment for hypertension, a condition that each year kills 10 million people around the world. 

Today, masked hypertension is detected with ambulatory blood pressure monitors, which take readings as a patient goes about their normal day. The equipment, however, can be cumbersome and expensive, and it may not be available in low-income countries. Doctors may also not even consider an ambulatory monitor for patients with masked hypertension who have normal blood pressure readings during an exam. 

The diagnostic tool created by the U of A reviews a range of health indicators to predict masked hypertension with 83% accuracy. As the tool trains on more data, its accuracy should increase. The new diagnostic tool also has a low rate of false positives, which would help patients avoid unnecessary and potentially expensive treatments. 

Examining the Data

The U of A researchers trained their tool to diagnose masked hypertension using African-PREDICT, a study that collected health data on 1,200 young people in South Africa. The data set was unique because it measured healthy men and women in their 20s and 30s from different ethnicities and a variety of socioeconomic levels. 

"As modelers, we're very data hungry. We're always looking around for what other people have measured experimentally," Richardson said. 

Richardson read about the study, emailed the organizers and started an international collaboration between Arkansas and South Africa. 

The South African study participants wore ambulatory blood pressure monitors, which identified who had masked hypertension.

The researchers asked the computer to look at two groups of participants: one with masked hypertension and another with normal blood pressure. Then, using machine learning, the computer examined the health and demographic data collected by the study to discover what the members of each group had in common.

"It's essentially pattern recognition," Richardson said. 

The computer can also make connections between data that humans might never see. 

"There's so much we don't know about biology. There is no human on the planet who knows how all those molecules work together," he said. 

To test the accuracy of the model, the researchers had the computer review another group of participants, but this time it was not told who had been diagnosed with masked hypertension. The computer had to make that diagnosis, which it did correctly 83% of the time. In the future, the model could be trained on other patients, which would help it refine the factors that indicate masked hypertension and become more accurate. 

The South African researchers are working on a 10-year follow-up to their study, which will also provide more data to refine the model. 

The Next Step

Artificial intelligence found the connections and created the diagnostic tool, but what doctors will use to diagnose masked hypertension will be a formula. The approach is fundamentally no different than the way doctors look at the ratio of "good" and "bad" cholesterol to estimate the risk of cardiovascular disease. 

"All our model does is take that math, the ratio of two different numbers, and does that same thing after measuring 100 different molecules in your bloodstream," Richardson said. 

The model also considers factors such as age, sex, race and other health information like heart size measured by ultrasound. 

The diagnostic tool would be integrated into medical record systems to automatically diagnose masked hypertension risk based on the data collected about a patient. Once put into clinical practice, the model could continue training on data from actual patients to increase its accuracy. 

Richardson hopes that AI tools will help save lives. He also hopes these tools can help doctors take better care of their patients. 

"How can we produce tools to do some of the robotic tasks for clinicians, so they have more time to have human interactions with a patient?" he asked. 

The results of the work were published in the journal Frontiers in Physiology. The other authors are Samuel J. Coeyman, also of the U of A, Brendyn Miller of Wake Forest University and Annemarie Wentzel and Carina M.C. Mells of South Africa's North-West University.