Finding the Balance
Pitt Researchers Use Bridges to Optimize Doctor Visits for Kidney Patients
The biggest causes of death in developed countries—heart disease, cancer, chronic organ failure—are complex conditions that may be best treated by long-term management rather than attempting to “cure” them. Researchers at the University of Pittsburgh have used the Bridges system at PSC to mine a massive database of medical records to optimize preventive doctor visits for individual patients with chronic kidney disease.
Why It’s Important
The biggest causes of death in developed countries—heart disease, cancer, chronic organ failure—are complex conditions that develop over decades from many different causes. It’s unlikely these diseases will yield to any single, simple “cure.” Prevention, both against developing these conditions in the first place and stopping them from progressing in people who do have them, seems more realistic.
But because of the complexities of the diseases, individual differences in patients can call for very different treatments, even when their stage of disease is similar. That’s why Zlatana Dobrilova Nenova of the University of Denver, then a graduate student working with Jennifer Shang at the University of Pittsburgh’s Katz Graduate School of Business, wanted to learn whether supercomputing could offer better recommendations for individual patient care to doctors treating people with chronic kidney disease (CKD), who account for about one fifth of all Medicare spending. Too few visits and doctors can miss a preventable worsening of the disease; too many, and the health system becomes clogged with patients who don’t need it yet, making care more expensive for everybody.
“The idea was … can we improve the treatment of patients with chronic conditions by optimizing the number of times they should see a doctor?”—Zlatana Dobrilova Nenova, University of Denver
How PSC Helped
Previously, researchers had worried most about when patients with CKD would progress to “end-stage” disease, which requires kidney dialysis. That’s hardly surprising: this end state is expensive to treat, hard on patients, makes other conditions worse and so is life-threatening. But Nenova, consulting with CKD-specialist physician John Hotchkiss at the Pittsburgh VA Hospital and the University of Pittsburgh Medical Center, wondered if the progression could be halted or at least slowed far earlier—and whether an improved appointment schedule based on individual patients might improve care more than the current “one size fits all” federal guidelines. These guidelines match up patients with different appointment schedules (seen every 12, 6, 3, or 1.5 months) based on their CKD disease stage, but no other information.
“If two patients are very similar in terms of their [medical] history, then it’s very likely that they are going to progress in the same manner.”—Zlatana Dobrilova Nenova, University of Denver
With the help of PSC’s Rick Costa, the researchers ran their “case-based reasoning” model for predicting the likely course of individual patients’ cases on the Bridges supercomputer at PSC. Nenova then fed these predictions into a simpler model, run on a normal computer, which calculated the optimal appointment schedule using the federal guidelines as a starting point. One surprising result was that the patients with the most advanced CKD did not necessarily rate the most frequent appointments. In retrospect, Nenova suspects, this reflects the fact that these patients have progressed to the point at which preventive appointments would no longer help them. They need to go to dialysis. In a more expected result, the models did recommend more frequent visits for patients with less advanced CKD but who had multiple other health conditions. Another application of the work is that a given clinic can use the models to tell, based on their total patient population, how many appointments their patients are likely to need. This in turn may give warning of when clinics need to hire more people. Future work may focus on specifying which type of appointment—for example, cardiologist vs. kidney specialist—each patient needs the most.
“In the models, you’re looking at a huge amount of data and pulling only a couple of instances that were similar enough to your new individual; that was very computationally intensive … I tried to use my laptop for it at first and it … would literally crash … At that point, my advisor got an email from PSC … and using your resources I was able to speed the thing enough that I could graduate in four years. I probably would have stayed there for another two years if I’d had to use even a very good desktop [computer].”—Zlatana Dobrilova Nenova, University of Denver
Deeper Dive: Case-Based Reasoning
To predict patients’ likely futures, Nenova developed a “case-based reasoning” model. This model looked at the entirety of each patient’s medical record over the previous 12 months, comparing it to a database of 215,821 patient cases. The idea was to match the patient being treated to others who were most similar over 12 months, looking at how the others progressed in the 13th month to predict how that month would go for the patient. One problem the researchers had to face was that they didn’t know ahead of time which medical data would be important, so they wanted to look at everything in the database. The flip side of that was the risk of irrelevant information overwhelming a few, more important, data points. Drawing on Hotchkiss’s clinical experience, Nenova’s computer model focused on signs and symptoms that showed the fastest, biggest changes before patients in the database grew sicker. Another issue was deciding how many similar people they needed to compare to the patient. Was the most similar case to the patient’s enough to predict his or her progression, or did comparing to more than one person improve the prediction? And if you made more than one comparison, should the model pay more attention to the cases most like the patient than to those a little less like the patient?
The final model, which balanced all these concerns in mining the massive database, proved too complex to run on a laptop or desktop computer. Applying Bridges’ raw power to that model, which was expressed in the statistical package R just as it was on her desktop, provided predictions for each patient’s likely progression.