Building Machine-Learning Algorithms that Reduce Hospitalizations and Costs

A researcher says that developing the algorithms is less of a challenge than the necessary investments in collecting data and installing electronic health records.

Healthcare analytics are frequently in the news these days, but what impact do they have on individual hospitals and health systems? The connection is much clearer at Boston University’s Center for Information and Systems Engineering, where researchers are bringing the power of machine-learning algorithms to bear on the problem of unnecessary hospitalizations of patients with heart disease and diabetes.

Drawing on anonymized EHR data, the Boston University researchers have been able to predict hospitalizations in this population a year in advance with up to 82 percent accuracy, theoretically giving hospitals the opportunity to intervene and treat the diseases more aggressively in an outpatient setting.

Leading the team is the center’s director, data scientist, and engineer Ioannis Paschalidis, a professor in Boston University’s departments of electrical and computer engineering, systems engineering, and biomedical engineering, and head of the Network Optimization & Control (NOC) Lab.

The focus of your work at NOC has been optimizing the design and operation of networks, as well as designing control algorithms to regulate their operation. How did you decide to focus on health care?

Paschalidis: I had read about the U.S. healthcare system’s attempt to become more efficient, reduce costs, and improve patient outcomes when I came across the AHRQ [Agency for Healthcare Research and Quality] study showing that we spent about $30.8 billion on unnecessary hospitalizations in 2006 (Jiang, J., Russo, A, Barrett, “Nationwide Frequency and Costs of Potentially Preventable Hospitalizations, 2006,” AHRQ, April 2009). I thought that people like me with a quantitative, optimization-oriented background could contribute something. We decided to start with heart disease, which alone accounted for almost a third of that amount, and later added diabetes, another huge part of the cost equation.

We set out to build software that could automatically flag patients at increased risk for medical emergencies based on data in their EHRs. It works by sifting through records of patients who were previously hospitalized and learning which risk factors—a certain number of chest complaints or an unusual level of a particular enzyme in the heart, for example—might have served as early alerts. We have also done some work on predicting 30-day readmissions following surgery; the hope is to guide postoperative care to prevent these readmissions.

We have established relationships with some of the local hospitals like Boston Medical Center and Brigham and Women’s Hospital to get data, and we’re planning to continue with this initiative on a long-term basis. The time has come to use data to help physicians make decisions more systematically and also to work through problems they cannot possibly solve themselves. They cannot check every patient they have ever seen to make these sorts of inferences, and neither can the hospital.

How do the accuracy rates of your predictions compare with other risk-scoring systems?

Paschalidis: They substantially surpass them. For example, the Framingham Heart Study’s 10-year cardiovascular risk score can predict hospitalizations with 56 percent accuracy. Even feeding the factors used in the Framingham score—such as age, cholesterol, weight, and blood pressure—into more sophisticated machine-learning methods leads to inferior results: 69 percent for the Framingham risk score compared with our 82 percent for our algorithm. This suggests that using the entirety of patients’ medical records, which can contain up to 200 factors, yields superior results.

Is this software currently in use at any hospitals?

Paschalidis: No, but we are talking with Boston Medical Center to see how these algorithms could be used as part of their regular clinical practice. Leaders there are interested in implementing the system so it will issue alerts for patients at higher risk for readmissions. That way, they could put protocols in place to pay more attention to those patients during and after their stays.

What has to happen for hospitals to be able to use this technology? What are the barriers?

Paschalidis: I don’t think there are big obstacles in terms of technology and tools; they already have the information in their EHRs. I think it’s more a matter of the hospital making the decision to invest the time, and some resources, to make this happen. Collecting the data is expensive and installing the EHR is expensive, but after that, it doesn’t take much to develop an algorithm that pulls the correct data and gives you an indication.

The problem is that without proper incentives, it’s not clear what the hospital would gain either financially or in terms of reputation. Patients benefit, and the U.S. economy gains from reducing unnecessary hospitalizations, but it’s not a huge item on the agenda for most hospitals. Unless their readmission statistics are below the national average, they’re not being penalized, and that is only a very recent development.

Then what will drive the adoption of the software?

Paschalidis: It depends on the business model of the hospital. I think hospitals involved in ACOs [accountable care organizations], like Boston Medical Center, will be the most interested because they share some of the risk with Medicare. If they can predict hospitalizations for patients with heart disease, then they can get a handle on their costs for caring for those patients for the next year. In life and in business, being able to predict the future is very powerful.

I expect the technology could be in place widely in three or four years.

Will hospitals draw on just their own data, or would the data be centralized regionally or nationally?

Paschalidis: There is the potential for a centralized function for hospitals that use the same EHR system, in which the algorithms are informed by data other hospitals have generated. This would be more helpful for smaller hospitals that don’t have enough patient data to feed the algorithms.

But my prediction is that in general, we will have a very distributed way of deploying these types of algorithms; it could eventually become another application that gets embedded in the EHR system, another data field. How hospitals use it and how they fine-tune the calculations will probably be a local matter.

You have said that hospitals are already making wide use of analytics but that we are only seeing the tip of the iceberg. What are some examples of current uses and what might we find beneath the waterline?

Paschalidis: Hospitals are adopting business analytics widely used in the transportation industry to schedule operating rooms and staffing. Other algorithms are being developed to help physicians make diagnoses. Our Boston University team has developed methods to automatically titrate medications in ICUs in response to the patient’s condition.

Eventually, I can see integrating data from myriad technologies—implantable devices such as pacemakers, fit trackers, and smartphones, even credit card and electronic payment systems—with the EHR to create a rich personal health record we can carry in our pockets.

Lauren Phillips is president of Phillips Medical Writers, Ltd., Bellingham, Wash..

Interviewed for this article:

Ioannis Paschalidis, PhD, is professor, Department of Electrical and Computer Engineering, Boston University College of Engineering, Boston.