Doctor, Doctor, Give Me the News: To Predict a Bad Case, Ask a GPU

by Jamie Beckett

Deep learning is giving doctors a life-saving edge by identifying high-risk patients before diseases are typically diagnosed.

Researchers at New York’s Icahn School of Medicine at Mount Sinai are using deep learning, a branch of artificial intelligence, to analyze patient electronic health records and determine those most likely to develop a serious ailment within the next year.

The experimental tool, dubbed Deep Patient, trained on 12 years of patient records – 700,000 in all. When tested, it could predict risk for dozens of diseases including heart failure, several types of cancer and severe diabetes.

“For most diseases, prevention is easier than reversal,” said Joel T. Dudley, an author on a recent paper in Nature Scientific Reports describing the Mount Sinai study. “This could have a huge impact on people’s health.”

Electronic Health Records Hold Promise

With the advance warning, doctors and patients gain precious time to take action that might prevent the disease or at least delay its onset, said Dudley, who is also an associate professor at Mount Sinai. The doctor could, for example, recommend a new medication or put the patient on a special diet.

Deep Patient, which is GPU-accelerated, could also save on healthcare costs, Dudley said. By focusing resources on preventing disease in high-risk patients, healthcare providers can avoid the more expensive cost of treatment later.

The researchers’ impetus for Deep Patient came in part from their frustration with electronic health records. The records contain massive amounts of information about patients – lab tests, surgeries, medications, medical history and more. But, to date, doctors haven’t been able to use the data to improve diagnosis or treatment, Dudley said.

Doctor reviews electronic health records on a tablet computer.
A doctor reviews a patient’s electronic health records. Image courtesy of  the U.S. Department of Agriculture under Creative Commons License.

“Electronic medical records are used for billing and not as a tool for doctors,” Dudley said.

The researchers want to use the records and deep learning to advance precision medicine, an approach to disease prevention and treatment that’s customized for each patient.

The researchers trained their neural network on thousands of patient records using NVIDIA Tesla K80 GPU accelerators and our CUDA programming model. They tested their models on about 75,000 patients.

Predicting Multiple Diseases Rather Than One

The Mount Sinai researchers are not the first to use electronic health records and deep learning to predict disease risk. (See “How AI Can Predict Heart Failure Before It’s Diagnosed”). But unlike earlier approaches that focused on a single disease, theirs includes nearly 80 ailments. They were able to do this by creating a new way of representing patients’ medical data for computer analysis.

“The old way to do this was to create one-off specialized data representations for a specific disease, customized for every situation,” Dudley said. “Patients don’t often have just one disease – they have several.”

Instead, the researchers created one representation that encompasses all of a patient’s medical history, said Riccardo Miotto, a data scientist at Mount Sinai and lead author on the paper.

More work needs to be done before doctors can use Deep Patient to help patients. The Mount Sinai team next plans to add more types of data such as genetic information and family medical history to make better predictions.

More Data, Better Decisions

Dudley said he hopes to incorporate data beyond electronic health records to create a tool that doctors could use to make better diagnostic and treatment decisions.

“One of the challenges is that physicians have less time to deal with patients and more information in electronic health records,” he said.

As more genetic information and data from wearable medical devices becomes available, “doctors are going to get inundated,” he said. “We want to build a system that uses all the data available to make predictions about patient health and provide new insights for physicians.”