Change of Heart: How AI Can Predict Cardiac Failure Before It’s Diagnosed

by Jamie Beckett

The last place you want to learn you have heart failure is where it often winds up being diagnosed: in the emergency room.

Researchers analyzing electronic health records are using  artificial intelligence and GPUs to get ahead of this curve. They’ve shown they can predict heart failure as much as nine months before doctors can now deliver the diagnosis.

A research team from Sutter Health, a Northern California not-for-profit health system, and the Georgia Institute of Technology, believe their method has the potential to reduce heart failure rates and possibly save lives.

“The earlier we can detect the disease, the more likely we can change health outcomes for people and improve their quality of life,” said Andy Schuetz, a senior data scientist at Sutter Health and an author of a paper describing one aspect of the research. “That’s what’s exciting to me – the potential to change the future.”

A paper on the complete research process and findings is being considered for publication by a major medical journal.

AI Gives Doctors an Edge

The stakes could hardly be higher. Each year, about 23 million people worldwide, including nearly 6 million Americans, have heart failure, according to the American Heart Association. It occurs when the heart muscle is weakened and can’t pump enough blood and oxygen to meet the body’s needs. Half of those diagnosed die within five years.

If doctors know which patients are likely to develop heart failure, they can prescribe medication or recommend lifestyle and diet changes that could delay its onset or even prevent it entirely, said Schuetz.

AI and NVIDIA are playing a growing role in advancing healthcare around the country. NVIDIA announced last week that it’s working with Massachusetts General Hospital to apply the latest AI techniques to improve disease detection, diagnosis, treatment and management.

Sutter Health logo The team analyzed electronic health records from more than 265,000 Sutter Health patients. From these, it studied 3,884 patients with heart failure and about 28,900 patients as a control group.

Researchers analyzed the records using deep learning, a type of artificial intelligence that can solve complex technical problems like face or speech recognition — sometimes even topping human performance.

This problem required computers to understand many types of health records — prescriptions or summaries of visits to doctors, for example — in many formats. Another challenge was tracking changes in the records over time to provide a full picture of patient health.

Deep Learning vs. Traditional Machine Learning

The problem of predicting heart failure onset isn’t new, said Jimeng Sun, an associate professor at Georgia Tech and an author of the paper. What’s different is deep learning and GPUs.

Unlike traditional machine learning, deep learning does not require a human expert to define every factor the computer should evaluate in the data — a time-intensive process. In earlier research, Sun said he and others spent a couple of years working with experts to build machine learning models.

Then Georgia Tech doctoral candidate Edward Choi, another author of the paper, spent a summer at Sutter Health and applied deep learning to the problem.

Tesla K80 GPU
Researchers used Tesla K80 GPUs to help predict heart disease risk.

“In three months, he was able to outperform what we’d done,” Sun said.

GPUs provided the speed required to train the neural networks — where the learning takes place in deep learning — on the hundreds of thousands of Sutter Health records, said Choi.

“Without widely available GPUs, this work would not have been feasible,” Schuetz added. The team used Theano, CUDA 7 and a Tesla K80 GPU in their research.

Method Applicable to Any Disease

Although this work focused on heart failure risk, the researchers say their approach could be used to predict any kind of disease.

Their next target may be detecting sepsis in emergency room patients, Schuetz said. Sepsis is a life-threatening response to an infection and the ninth-leading cause of death in the U.S. But it’s maddeningly difficult to detect. By the time doctors realize a patient has sepsis and start treatment, it may be too late.

Researchers are also working on a platform that will let them put the prediction tools into the hands of doctors, Schuetz said.