Deep Learning System Triages Terminally Ill Hospital Patients

Posted on January 26, 2018 I Written By

Anne Zieger is veteran healthcare branding and communications expert with more than 25 years of industry experience. and her commentaries have appeared in dozens of international business publications, including Forbes, Business Week and Information Week. She has also worked extensively healthcare and health IT organizations, including several Fortune 500 companies. She can be reached at @ziegerhealth or

Researchers at Stanford have developed a new tool designed to coordinate end-of-life care for critically ill patients. While the pilot study has generated screaming newspaper headlines (“AI tool predicts when people will die!”) researchers say that the system is best thought of as a triage option which helps hospitals and hospices provide timely palliative care to those who need it. It can also help terminally ill patients — most of whom would prefer to die at home — make plans for their passing and avoid dying in their hospital bed.

According to an article in tech publication Gizmodo, the Stanford set-up combines EHR data with other sources of information such disease type, disease state and severity of admission. The information is then processed by a form of AI known as deep learning, in which a neural network “learns” by digesting large amounts of data.

To conduct the study, researchers fed 2 million records from adult and child patients admitted to either Stanford Hospital or Lucile Packard Children’s Hospital. The system then identified 200,000 patients who met the study’s criteria. In addition to clinical criteria, the system also reviewed associated case reports diagnoses, number of scans ordered, number of procedures performed and other data.

After reviewing 160,000 case reports, the deep learning system was instructed to predict the mortality of a given patient within three to 12 months of a particular date using EHR data from the previous year. The algorithm included a requirement to ignore patients who appeared to have less than three months to live, as this window was too short for providers to make preparations to offer palliative care.

Then, the AI algorithm calculated the odds of patient death in the 3 to 12-month timespan extending from the original date. Its predictions turned out to be quite accurate. For one thing, it predicted patient mortality within the 3 to 12-month window accurately in nine out of 10 cases, a performance that few clinicians could match. Meanwhile, roughly 95% of patients considered to have a low probability of dying within 12 months actually lived beyond that point.

It’s worth noting that while the deep learning tool made fairly accurate predictions of patient mortality, the system doesn’t let healthcare providers know what treatment patients need or even how it makes its predictions. Luckily, researchers say, the system allows them to get a look at individual cases to better understand its deductions.

For example, in one case the system predicted accurately that a patient with bladder and prostate cancer would die within a few months. While there were many clues that he was near death, the system weighted the fact the scans were made of his spine and a catheter used in his spinal cord heavily in its calculations. Only later did the researchers realize that an MRI of the spinal cord most likely suggested a deadly cancer of the spinal cord which was likely to metastasize.

It’s worth remembering these results were produced as part of a pilot project, and that the predictions the system makes might not be as accurate for other data sets. However, these results are an intriguing reminder of the possibilities AI offers for hospitals.