An artificial intelligence program developed by researchers at Google can predict when a hospital patient will die with up to 95 percent accuracy, according to a new paper published in the journal npj Digital Medicine. If this type of AI is implemented in hospitals, it could help these hospitals save money and more efficiently apply their resources, increasing the number of lives they could save.
Google’s hospital AI uses the same technology behind many of the company’s other AI projects, like its image recognition software or DeepMind programs that can play Go and optimize server rooms. In this instance, Google trained its AI to evaluate patient conditions based on thousands of factors available in those patients’ health records.
Using that information, the software can predict the likelihood that a patient will die while in the hospital, and how long it will take for that to happen. This information can give doctors and other hospital staff valuable information about which patients to spend the most time treating and which patients are the most likely to successfully recover.
In a trial run of this new software at two U.S. hospitals, the Google AI was 95 percent and 93 percent accurate in its predictions of patient mortality. This is a dramatic improvement over traditional hospital software, which average about 85 percent accuracy. Primarily, this is due to the number of variables used by Google’s AI. The AI analyzes over 100,000 factors in order to make its predictions, compared with only a few dozen or less for most other models.
Machine-learning technology plays a significant role in making this type of analysis possible, but credit also goes to the recently introduced Fast Healthcare Interoperability Resources (FHIR) standard, which allows hospitals and other healthcare providers to share patient data in a much more accessible format. With this standard, the researchers could include thousands of bits of information, up to and including borderline unreadable freehand notes from doctors.
Source: npj Digital Medicine
