Google Trained Machines To Predict When a Patient Will Die

Henrietta Strickland
June 21, 2018

Patients who are at high risk of death due to preventable factors can be attended to.

The system is still in its infancy, but Google believes it could someday be used to predict death far longer in advance. In trials that used data from two US hospitals, the researchers were able to show that the algorithms in the model improved predictions as to the length of stay and time of discharge - but also the time of death of patients. This method was 85 percent accurate at the University of California, San Francisco system and 83 percent accurate at the University of Chicago Medicine system, the study says.

Additionally, the more efficient processing of data could also mean that health-care workers would spend less time on paperwork and focus more on patient wellbeing.

"'When will I be able to go home?" Nearly 80 percent of the time is spent on making the data presentable, said Nigam Shah, co-author of the Google's research paper and an associate professor at Stanford University. One problem is to amass all the data about patients in one central database as the data is often spread widely through various healthcare systems and government agencies often without the data being shared. No one can say for sure how much time you have left or what are your chances of beating the odds. Shah said, "You can throw in the kitchen sink and not have to worry about it".

'When will I be able to go home? Will I get better?

Google's predictions on all of these fronts are quite impressive. Specifically, the company's Medical Brain team developed an algorithm that can assess health risks, including how likely a patient is to die. Uses of the predictions The predictions may enable the hospital to prioritize patient care, or adjust treatment plans, and even recognize medical emergencies before they happen.

For each prediction, a deep learning model reads all the data-points in electronic health records, from earliest to most recent, and then learns which data helps to predict the outcome. In that particular case, the challenge comes from inaccuracy - even the smallest mistakes in a patient's record can result in them getting the wrong care. The research paper was published in the Nature Journal.

The team took a different approach to building predictive statistical models by considering a "representation" of all a patient's health records, including clinical notes, rather than removing most of a patient's information from the analysis.

The current method of collecting data is cumbersome, costly, and time-consuming.

The system is based on the development of neural networks suitable for Autonomous learning and development.

Other reports by Click Lancashire

Discuss This Article