The same machine learning technology Google uses to anticipate the next word in a search query is in tests to advance predictive modeling in health care, and the tech giant is teaming up with three universities this year to research its ability to do so.
Katherine Chou, head of product for Google’s efforts in machine learning for health, has been working on using machine learning for preventive care since 2015, when health care systems reached out to Google for help. Part of the breakthrough now is that “our deep learning technology is finally mature enough to handle the unstructured data that you see in the records,” she says.
An example of unstructured data is pictures and notes from a physician or nurse that make up the majority of an electronic health record. Google’s machine learning can now include those data in a predictive modeling algorithm.
“Here’s what’s exciting to me. Predictive modeling until now has effectively thrown away large amounts of data in the electronic health record because the statistic models can’t use the notes or X-rays or pictures,” says Michael Howell, chief quality officer at University of Chicago Medicine, one of the three academic institutions along with Stanford Medicine and the University of California, San Francisco that is collaborating with Google. “What they bring to the table is this amazing ability to deal with data that doesn’t fit into a spreadsheet.”
Want to publish your own articles on DistilINFO Publications?
Send us an email, we will get in touch with you.
Howell has been involved in predictive modeling for more than 15 years, when the problem was data scarcity, and modelers had to manually collect data and build models. Now, with robust electronic health records, the issue is having too much data. “More than nine out of 10 health care encounters in the United States leave a significant electronic footprint,” he says.
In fact, in 2008, only 1.5 percent of U.S. hospitals reported having a comprehensive electronic health record, according to Howell. That changed fast thanks in part to the Health Information Technology for Economic and Clinical Health Act, and, by 2015, 90 percent of hospitals qualified as having meaningful electronic records.
In addition to the ability to make predictions using unstructured data, Howell sees another key win with Google’s ability to create and organize data. He says the act of pulling data from a record or repository and the creation of a data set for a predictive algorithm is 90 percent of the work. “Google is obviously world-class in this preprocessing step of how to get the data in shape to do predictions,” he says.
Chou says that machine learning technology is helping to automate and harmonize the way data appear to different health care organizations, by working in concert with the open data Fast Healthcare Interoperability Resources standard. “Each clinic has its own way of recording data, which makes gathering insights from multiple clinics difficult.”
Google has seen its machine learning work to improve care already. In India, medical imaging predictions enabled Google to screen patients for diabetic retinopathy and see a doctor sooner than they would otherwise. The study was published in the Journal of the American Medical Association.
The University of Chicago has successfully implemented predictive modeling to reduce cardiac arrests in the hospital. Matthew Churpek, M.D., and Dana Edelson, M.D., developed an algorithm from vital signs, lab results and demographic data in an electronic health record that predict a patient’s risk of cardiac arrest. The two formed a company called Quant HC to bring the tech to other hospitals.
Named eCART, the tool has been layered into technology at the hospital so that a computer pages a nurse to respond to patients deemed high risk. “This is an important point: An algorithm by itself does nothing,” Howell says. “You need a response in health care to do something.”
Howell says the same goes for the Google partnership. He couldn’t discuss specific predictions that the research is looking at because the work is still in proof of concept and has been submitted for peer review.
Date:Aug 08, 2017