Each year, approximately 209,000 patients suffer from an in-hospital cardiac arrest in the United States, with less than a quarter living long enough to be discharged. While rates of survival have improved over the last decade or so — likely due to improvements in quality of care like quicker response times and well-trained personnel — what if cardiac arrests and other adverse in-hospital events could be prevented from happening altogether?
With the vast majority of hospitals now using electronic health records, big data researchers are aiming to use this new influx of information to predict and prevent patient deaths. So instead of responding to a cardiac arrest as it happens, their prediction models would pick out the patients most likely to have one so doctors can treat them accordingly and hopefully reduce the likelihood of future issues. Predictive modeling using big data is a hot-ticket strategy already in effect across a number of industries. Netflix has 800 engineers who work on the algorithms that power their personalized viewing suggestions, which forecast movies and shows you'll like based on your watch history and ratings. Credit scores use variables like number of late payments and credit card utilization to rank individuals based on who is most likely to default on a future debt.
One big data researcher who foresees a similar future in health care is pediatric oncologist Samuel Volchenboum, who directs the Center for Research Informatics at the University of Chicago. Volchenboum and his team of data experts have combed through countless electronic health records dating back to 2006 to create the Clinical Research Data Warehouse, an enormous repository of medical data meant for research purposes. We spoke to him about how the new age of big data and predictive algorithms could potentially make hospitals a safer place for patients.
Q: Can the large amounts of data being collected by hospitals be used to predict patient deaths before they happen?
Concurrent with the installation of electronic medical records at most hospitals now is our ability to collect an ever-increasing amount of data on these patients. Whereas research before had to be done through pretty painstaking mining of written notes, one is now able to extract data directly from the electronic medical records.
Once we've gone through the process of cleaning up the data to be sure we have high-quality information, we can then study some of these adverse events and what might cause them to occur. One of the first big studies we did at the University of Chicago using our Clinical Research Data Warehouse was by two of my collaborators, Matthew Churpek and Dana Edelson. They took data on nearly 60,000 admissions and built a mathematical model called eCART to predict patients most likely to suffer cardiac arrest on the hospital wards. The prediction model is based on things you'd expect like respiratory rate and blood pressure, but also lab values and many other inputs.
Q: How does the hospital use this prediction algorithm to help patients?
When a patient achieves a certain score according to this proven algorithm, he or she is at a much higher risk of having a cardiac arrest. So at our hospital now, there is a response team that actually monitors patients' scores that are being calculated in real time.
If the score goes above a certain threshold, the team will go to the patient's room and try to intervene or fix whatever factors are contributing to the patient's high level of risk. This system has been in use for several months and has been shown to be effective in alerting which patients are at the greatest risk of declining.
In the long run, hospitals and health-care providers looking at their systems for these kinds of trends is what will make patients safer, and so that's what I think is really encouraging. There are even groups working on “App Stores” for these prediction algorithms, where hospitals can download them to monitor their own patients.
Q: What other sorts of prediction algorithms are currently being developed or tested to prevent adverse outcomes?
I’m working with a group at the University of Chicago that is trying to build a sepsis algorithm that will help predict which patients will become infected in the ICU. Such algorithms are obvious applications of these data, but what I'm actually most interested in is trying to bring in factors that are less obvious yet may still influence patient outcomes.
For instance, how does crowding in the emergency room affect the care of patients up on the hospital floor? Or how does a nursing strike affect the length of stay for patients? Or even how does a storm in the forecast or traffic congestion affect the care of patients on the ward? These things would not necessarily seem connected, but as we collect more and more data, these sorts of links could appear.
To make hospitals safer, we need to look at many potential relationships between events. To do this, we need “big data” – the kind of data used for the eCART study above. By having large, well-curated data sets, we can now study highly-complex relationships between seemingly unrelated events. We’re even in a position to look at how the care of one patient might affect patients around them. With so many adverse events affecting care, it follows that we should be looking at the downstream impacts of adverse events.
For more health news, you can sign up for our weekly newsletter here.