Machine-learning algorithms are now being used to provide medical diagnoses, to manage stock markets, and to write newspaper stories. They can also be used to predict the future.
In a paper published in the journal Nature, researchers from the University of Washington and Google describe a new machine-learning algorithm that can predict the future of a person’s health. The algorithm was able to predict whether a person would be diagnosed with diabetes within five years with an accuracy of 81 percent. The algorithm was trained using data from electronic health records (EHRs) of 1.6 million patients. It was then tested on data from another 1.3 million patients.
“We have a lot of data about people, and we know how they are doing in the future,” says senior author Rajesh Rao, a UW professor of computer science and engineering. “Can we use that information to learn enough about the underlying biology of the disease that we can actually make predictions about the future?”
The researchers used machine-learning algorithms to analyze EHRs from patients who had been diagnosed with diabetes. The algorithm learned patterns in the data and was then able to predict whether a patient would be diagnosed with diabetes within five years. The researchers also tested the algorithm on patients who did not have diabetes but were at high risk for developing it.
The algorithm predicted which patients would develop diabetes with an accuracy of 81 percent.The researchers hope that their work will lead to better ways to prevent and treat diabetes. They plan to use their algorithm to identify new drug targets for diabetes and other diseases. They also want to explore using their approach for other diseases, such as heart disease, cancer, and Alzheimer’s disease.
“That’s important in diabetes because we don’t have a great way to attack it other than lifestyle changes and drugs like metformin,” says corresponding author Eric J. Topol, a geneticist who directs the Scripps Translational Science Institute in San Diego, Calif. “Drugs like metformin would be more effective if we used machine learning to make them individualized. You and I might respond differently to the same drug. If the drug companies knew how to use this approach, they could improve the efficacy of the drugs they already have.”
The algorithm was trained using data from people who had been diagnosed with diabetes. It was then tested on patients who were at high risk for the disease but had not yet been diagnosed.“We want to ask, ‘How could we get more healthy years from a patient who doesn’t have diabetes yet? How could we get that from someone already diagnosed?” Topol says.Diabetes is associated with heart disease, stroke, blindness, kidney failure, and limb amputation.
Most researchers believe diabetes is caused by a combination of genetics and environmental factors such as diet and lifestyle. However, there is still tremendous complexity surrounding the condition.“Think about figuring out what caused a car accident, not just for legal reasons but to make sure you prevent that type of accident from happening again,” Topol says. “We need a similar approach for diabetes. Even if we understood all the genetics, we still wouldn’t be able to predict what will happen with any individual person. Somebody might inherit many diabetes-associated genes and early trajectories and then at some point one bad thing happens and boom — diabetes. We have a 15-year timeline to understand this biology before kids get it. We have to change our basic model from understanding individual genes to studying populations in a way that allows us to use technology to affect behavior.”
Topol envisions doctors using personal data collected from wearable devices to make treatment decisions. Patients might one day wear glucose-sensing contact lenses to collect both medical and environmental data.
“Once we have heads-up displays in our glasses or smartwatches like a Star Trek communicator, we can do that easily,” Topol says. “You would be standing in line for the latest iPhone, and the kid in front of you would be getting a glucose monitor prescribed by his doctor based on his location alone. We’ve made a huge amount of progress in AI in the past six months. All of this sounds like science fiction now, but in five to 10 years, to be without this technology will seem crazy.”
Topol also envisions the Algorithm predicting what personal decisions a person might make in terms of their health to either treat or delay disease. “We are teaching machines how to schedule your life by teaching it something like wisdom,” says Topol.
The researchers are exploring other uses for their machine-learning algorithm. “Tracking historical data allows us to see many patterns we had never seen before, Rao says. “For the first time, everyone has digital footprints that we can study. The hope is that eventually it will give us more information about these predictive signatures and ultimately help people live better lives. It could be used on other environmental exposures and be tailored to specific populations based on age and location.”
This article was written by AI, its was neither proofed nor researched.