Google says its deep learning model bests traditional methods in sorting EHR data
- A Google deep learning model outperformed traditional methods of sifting through voluminous EHR data in predicting health-related outcomes, according to a new study in the journal Nature. The paper was published in conjunction with this week's Google I/O conference.
- The internet giant worked with the University of California, San Francisco, Stanford University and the University of Chicago Medicine to assess the accuracy of deep learning and FHIR, a standard for easing the flow of data across disparate systems, in predicting in-hospital deaths, 30-day readmissions and long length of stay. In all three cases, the test model bested other EHR models.
- One impact from the improvement was a 50% reduction in the number of alerts needed to assess patient mortality, the authors note, leading to far fewer false positives.
The study included EHR data from 216,221 adult patients hospitalized for at least 24 hours. The combined data points totaled about 47 billion.
“We hypothesized that these techniques would translate well to healthcare; specifically, deep learning approaches could incorporate the entire EHR, including free-text notes, to produce predictions for a wide range of clinical problems and outcomes that outperform state-of-the-art traditional predictive models,” the authors say. “Our central insight was that rather than explicitly harmonizing EHR data, mapping it into a highly curated set of structured predictors variables and then feeding those variables into a statistical model, we could instead learn to simultaneously harmonize inputs and predict medical events through direct feature learning.”
AI is hot in healthcare right now and Google is eager to claim a share of the space. Its parent company Alphabet is casting a wide net over healthcare, but is positioning itself as an AI/machine learning company, according to a recent report from CB Insights. The analysis noted Google is looking to power a healthcare data infrastructure layer, which would use Google Cloud and include new data pipes for providers and payers.
Earlier this month, Fitbit announced it will combine its recently acquired Twine Health platform with Google Cloud's Healthcare API to connect user data and individual health records, with the goal of helping physicians and patients better manage chronic conditions.
In December, the tech giant launched Deep Variant, an open-source tool that uses AI to create an image of a person’s genetic blueprint using sequencing data. The aim is to home in on genes or gene mutations that help guide a patient’s diagnosis and treatment. The company’s DeepMind is also mining Big Data via a partnership with Moorfields Eye Hospital NHS Foundation Trust in the UK. That effort is looking at whether machine learning technology can effectively analyze ocular scans, in hopes of enabling earlier detection and treatment of degenerative eye diseases.
A study earlier this year in Nature Biomedical Engineering also reported on research by Google parent Alphabet and its research arm Verily Life Sciences that showed a way to predict a person’s risk of a major cardiac event using eye scans and deep learning.
- Nature Scalable and accurate deep learning with electronic health records
- Becker's Health IT & CIO Report How Google is using deep learning to understand EHR data and reduce readmissions