blood pressure monitoring equipment on a computer

Machine learning, a branch of computer artificial intelligence, has been put to use in a range of healthcare fields, from identifying promising drugs to recognizing changes in medical images that can catch health issues early. 

A new project at the College of Public Health aims to apply machine learning techniques to the everyday health of millions of people who have hypertension, a main risk factor for cardiovascular disease (CVD). Gabriel Tajeu, assistant professor of Health Services Administration and Policy, has been awarded a five-year, $748,000 grant from the National Institutes of Health to build a machine learning system to analyze electronic health data, identifying patients who are likely to have persistently uncontrolled blood pressure (BP) so doctors can be more proactive with their interventions.

Adults who have controlled BP have a substantially decreased risk of CVD-related mortality compared to those with uncontrolled BP. However, using current guidelines, only 40% of the 100 million U.S. adults with hypertension have controlled BP, making it a significant public health issue and one that is insufficiently kept in check. In treatment, there can be what’s called “clinical inertia,” Tajeu said.

“Even if a patient’s blood pressure is uncontrolled, the doctor might take a wait-and-see approach rather than increase the dosage of antihypertensive medications,” he said. “But six months or a year is a long time to be walking around with uncontrolled hypertension.”

What Tajeu’s machine-learning algorithm would do is provide the clinician with a prediction, based on a mix of data factors, that a particular patient is more likely to come back to their next appointment still having uncontrolled hypertension. “Hopefully that would give them more of a justification for increasing treatment intensity and lower that clinical inertia,” he said.

Machine learning algorithms have been put to work on electronic health records in applications such as predicting lengths of hospital stays and hospital mortality rates, but analysis of health data to identify trends and make predictions around hypertension is relatively new territory, Tajeu said.

A machine-learning algorithm (MLA) needs to be “trained” using a large set of data. In this case it will be anonymized data from the Temple Health System, which contains extensive demographic, clinical, prescribing, and dispensing data. An MLA can recognize patterns in data that may be difficult to detect otherwise and identify variables that are important for prediction of an outcome. Essentially, the system will assess existing data to analyze what combination of factors are most likely to indicate that a patient with uncontrolled BP will later return still having uncontrolled BP.  

Tajeu will work with graduate students in public health and computer science to build the system. By the end of the research, they hope to have a validated machine-learning algorithm. The next step would be to identify how to best integrate it into clinical management software used in the Temple Health System. That would allow a physician treating a patient to be alerted that a patient is at risk at the point of service, so appropriate measures could be taken. 

“This also allows you to use your resources in a targeted fashion,” Tajeu said. “Because if you just say for everybody who has uncontrolled blood pressure at a visit, we're going to do X, Y, and Z, that's expensive. But if there's a segment of that population that the machine learning algorithm identifies as the most high-risk, then we can target those patients. And in that way, we can save the healthcare system money, we can improve care for a vulnerable population in North Philadelphia, and ultimately save lives.”