Prevention by Prediction: How AI Spots Early Warnings of Disease

by Jamie Beckett

Eat healthy. Sleep enough. Exercise regularly. And ask your doctor about AI.

AI could soon help you stay healthy longer. Thanks to GPUs and deep learning, physicians can predict the onset of diseases far earlier than is now possible, simply by analyzing patient electronic health records (EHRs).

“We’re moving from treatment to prevention,” said Narges Razavian, a professor at New York University’s Langone School of Medicine. “We want to know, is this person at risk for something and can we predict it?”

AI to Predict Disease

In a talk at this month’s GPU Technology Conference, Razavian explained how her NYU team predicted 200 ailments three months faster than traditional methods by analyzing EHRs such as lab tests, doctors’ notes and X-rays.

Razavian’s deep learning software accurately predicted heart failure, severe kidney disease, liver problems, diabetes and hormone-related conditions based on just 18 common lab measurements captured over three years.

“Lots of diseases are preventable, but they happen so slowly that people get worse without realizing it,” Razavian said. “If we can use deep learning as a powerful tool to give patients a wake-up call, we’d be able to prevent diseases when there’s still time.”

By using deep learning to predict disease, researchers hope to help patients avoid cumbersome treatments like kidney dialysis.
The NYU team predicted severe kidney disease with its deep learning software. Detecting kidney disease early could keep patients off dialysis.

Researchers Find Hidden Links to Disease

The NYU researchers aren’t the first to realize the potential of EHRs to keep people healthier longer. (For examples, see “Doctor, Doctor, Give Me the News” and “How AI Can Predict Heart Failure Before It’s Diagnosed.”)

What’s different is the team’s ability to combine many different types of records over time to find previously unknown relationships, Razavian said.

For example, researchers predicted Type 2 diabetes by drawing connections among 900 measurements such as the patient’s weight, blood pressure, glucose levels, liver function and cholesterol levels. In the process, they found that some factors not typically associated with diabetes – a history of sleep apnea or acute bronchitis, hypothyroidism and anemia – may also predict the disease.

“This gets us closer to the biological mechanisms of diseases,” Razavian said.

How Deep Learning Aids Prevention

To predict disease, the researchers trained two neural networks on lab measurements and diagnosis information for 200,000 people selected from among 4.1 million insurance subscribers. The data was “raw,” meaning it had not been labeled or processed.

The team put its research to the test to improve an NYU medical center initiative that provides nurses’ visits and phone calls for 250,000 high-risk patients. By using deep learning to predict which patients were likely to suffer certain conditions, they helped the hospital determine who might be helped by an intensive lifestyle-management program aimed at preventing disease.

The work also helped automate scheduling for high-risk patients’ nurse visits and screening tests.

Razavian and other researchers hope to use EHRs and deep learning to advance precision medicine, an approach to disease prevention and treatment that’s customized for each patient.

“The applications for this work are enormous, and we’re just limited by personnel and time,” she said.