Using AI and Deep Learning to Predict Mortality
A big data and artificial intelligence software model predicts whether hospitalized patients will survive, better than traditional models.
Google, United States
"In this world nothing can be said to be certain, except death and taxes," wrote Benjamin Franklin. While we will all one day shuffle off this mortal coil the timing and nature is not certain. However, a new deep learning approach that incorporates big data from electronic health records (EHRs) is able to provide some surety by being able to predict when someone's time is up with 95% accuracy. That's according to a paper (Scalable and Accurate Deep Learning with Electronic Health Records) by scientists at Google.
Predicting Patient Outcomes
Computer scientists and medical experts from Google as well as three US universities claim their AI-based software can predict hospitalized patient outcomes better than traditional predictive models.
To make its predictions Google's model used more than 46 billion individual data points drawn from the electronic health records (EHRs) of over 216,000 patients in two hospitals. The medical-records data included patient demographics, laboratory results, previous diagnoses and procedures and vital signs.
Developing predictor models is challenging because there are thousands of potential predictor variables. Traditional models rely on the time-consuming task of manually selecting the relevant variables to use. Google's new model could make that process a lot easier.
“For each prediction, a deep learning model reads all the data-points from earliest to most recent and then learns which data helps predict the outcome,” said lvin Rajkomar MD, Research Scientist and Eyal Oren PhD, Product Manager, Google AI, in an interview with healthanalytics.com.
The biggest claim made by the research paper is that the software model can predict patient deaths 24-48 hours before current methods. This would allow doctors to administer potentially life-saving interventions. The model was also able to predict death at 24 hours after admission with 93% to 95% accuracy. This is around 10% better than traditional models.
Beyond being able to predict someone's expiry date with reasonable accuracy the AI-based software model was also able to predict unplanned re-admissions to medical facilities within 30 days as well as probable length of stay and diagnoses.
Limitations and Potential
The scientists involved in the research noted that there are limitations with their model. Among them that it was retrospective and that it hasn't been demonstrated that the approach has improved the delivery of care to patients.
However, what it has shown is the potential for an AI, deep learning and big data approach to understand better and improve healthcare delivery. In the Google paper, researchers wrote: “It is widely suspected that use of such efforts, if successful, could provide major benefits not only for patient safety and quality but also in reducing healthcare costs.”
Next Story »