In the newly released, non-peer-reviewed research paper, study investigators at the tech giant used de-identified data from collected from 216,221 patients over 11 years from both the University of California San Francisco Medical Center and the University of Chicago Medicine hospital to create a system looking to predict medical outcomes for hospital patients.
Google claims that the results, which have not yet been validated by independent sources, showed significant improvements over traditional models, including the ability to predict patient deaths one to two days before current methods are able to, according to the study.
Study researchers said that training AI to interpret vastly different handwriting, note-taking formats and what often seems to be non-sequitur data on electronic health records has been a major hurdle for such predictive systems.
To overcome this, Google researchers claim to have used three deep neural networks to sort through data and identify which pieces had the most impact on patient outcomes, Quartz reports.
The researchers also used a previous Google project, known as Vizier, to automatically train the system on how to interpret the data after it was ingested.
The project itself shows that Google is investing a significant amount of work into applying its artificial intelligence system into healthcare fields outside of its established healthcare plays, like parent company Alphabet‘s Verily.
Last September, Verily was reported to be developing a new artificial-intelligence powered test that searches for indicators of heart disease risk present in retina images.