Predicting death may soon be a feat of more than 90 percent accuracy for clinicians equipped with Google’s newly developed AI network.

The multinational tech giant is leveraging the potentials of machine learning and raw EMR data to help predict the course of disease and risk of death during hospital stays for patients, according to Bloomberg News.

Developing their approach at the University of California, San Francisco and the University of Chicago, researchers trained deep learning models on more than 216,000 deidentified EMRs from over 114,000 adult patients hospitalized at either location for at least one day.

Training focused on risk of mortality, readmission, prolonged stays and discharge diagnoses based on ICD-9 code standards with the algorithm incorporating the entire EMR, including free-text clinical notes and more than 46 billion pieces of other less-structured data for predictions made at discharge.

Information collected was then processed and compiled into Fast Healthcare Interoperability Resources (FHIR), an EMR data structure which is more flexible than the routine processing of mapping EMR data onto variables in statistical models.

The network established a 95 percent accuracy rate in predicting the risk of death in patients, and was encumbered by a lower number of false alerts compared to the augmented Early Warning score which assesses 28 factors and was 85 percent accurate at both locations.

The use of FHIR enables the data to be deployed to new facilities with relative ease compared to conventional predictive models, in which 80 percent of its work ethic focuses on organizing data to be presentable, although FHIR models trained on a particular site’s information may not be immediately transferable to a new site.

In addition, the solution informs clinicians of the sources from which data is pulled, including parts of a patient’s medical history, radiology findings or provider notes, decreasing any discomfort in relying on a "black box" of a neural network to produce a diagnosis.

First announced in May 2017, the endeavor has made researchers more confident in the development of accurate predictions that may lead to lower healthcare costs and fewer false alerts that hold potential for reducing alarm fatigue by physicians and nurses. More research and prospective trials are required to demonstrate its effectiveness and scalability.

“As with traditional tools and systems, the outcome measures covered in the research go beyond in-patient mortality, including length of stay, readmission, discharge diagnosis,” Iz Conroy, director of communications for health at Google, told HCB News.

The findings are available in npj Digital Medicine.