UD scientists merge electronic health records to improve care, clinical outcomes for CKD patients

Electronic health records have the potential to reduce the incidence of medical errors, duplication of tests, and delays in treatment.

But problems can arise when health care providers use different EHR systems that don't "talk to each other," and these problems are particularly critical for patients with multiple chronic diseases, or co-morbidities.

Now, a team of physicians from Christiana Care Health System and computer scientists from the University of Delaware is using merged electronic health records from various institutions to improve coordination of care and clinical outcomes for patients with chronic kidney disease (CKD), a condition that affects some 26 million American adults.

The team is led by nephrologist Claudine Jurkovitz, physician scientist in the Value Institute at Christiana Care, and Hagit Shatkay, associate professor in the Department of Computer and Information Sciences at UD.

With support from the Delaware ACCEL program, they are using longitudinal data collected from a large pool of patients to make predictions, initially about hospitalization patterns and later about other trends in the disease.

"Hospitalizations increase in frequency as kidney function declines and are mostly due to cardiovascular events or infections," says Jurkovitz. "We'd like to be able to predict the risk of hospitalization within a defined time period following an office visit."

"Our hypothesis is that, most often, catastrophic events leading to hospitalization are preceded by the convergence of trends such as increases in blood pressure, fluid overload, and the frequency of skin or respiratory infections, as well as uncontrolled high blood sugar for diabetic patients and declines in natremia, which is a drop in sodium levels in the blood."

This is where the expertise of the computer scientists comes in.

Shatkay explains that a probabilistic model for hospitalization, developed based on patients' longitudinal data, reflects a large number of variables, including trends in blood pressure and laboratory results, changes in medications, and frequency of outpatient visits and phone calls.

"These rich and diverse data require that we develop and examine machine-learning based methods for representation of, and prediction from, such data," she says.

Basically, the strategy she and doctoral student Moumita Bhattacharya are using involves identifying the features that carry the most information about the predicted target — in this case, hospitalization — so that the data can be represented more compactly. They are also focusing on the inter-dependencies among the various measurements so that predictions can be made more efficiently.

The next step in this phase of the research will be to examine the care of a cohort of children who transitioned from pediatric to adult care at Nemours/Alfred I. duPont Hospital for Children over a recent seven-year period.

Specifically, the researchers want to assess control of cardiovascular risk factors and long-term outcomes, as well as to acquire new information regarding medication adherence. Their hypothesis is that the latter is a major determinant of clinical outcomes in young patients with CKD transitioning to adult care.

Shatkay emphasizes that while this research focuses specifically on the effectiveness of the health care delivery system in providing and coordinating care for patients with CKD in Delaware, the approach the team is developing will have broader applicability — demographically, medically and geographically.

"The number of patients is in the thousands, and each one is associated with about 100 pieces of information," she says. "We're figuring out ways to sift through these massive amounts of data and determine what's relevant and what's not."

For Jurkovitz, the reward will be better care coordination for all CKD patients, with fewer of them being hospitalized.

"The rate of hospitalization for patients with CKD is as high as 430 per 1,000 patient-years," she says. "If we can identify the patients most at risk of being hospitalized, we can bring this number down. But first we have to figure out which factors are the most predictive of the decline that leads to hospitalization."