When we think about big data, we often picture it being used to address new challenges such as pharmacogenomics. It’s nice to remember that there are big data efforts tied to something relatively simple, like making sure patients get their blood pressure taken.
Last week I had a conversation with John Cuddeback, M.D., Ph.D., chief medical informatics officer for AMGA Analytics, a subsidiary of the American Medical Group Association. He described the process involved in combing through claims and EHR data to find patients with undiagnosed hypertension who are “hiding in plain sight.”
AMGA’s Analytics for Improvement (A4I) is a learning collaborative of its member practices working on population health improvement. The providers use the Optum One analytics platform. Leveraging the data, AMGA Analytics describes its work as creating analyses looking at the relationships among care processes, clinical outcomes and cost.
In its work on catching undiagnosed hypertension, AMGA Analytics took advantage of a prevalence estimator created by the Centers for Disease Control. Using the National Health and Nutrition Examination Survey (NHANES), CDC created a hypertension prevalence estimator that provider organizations and public health agencies can use if they know the demographics of their patient population, “The question for a provider organization is, ‘are we missing the diagnosis of hypertension on any of our patients?” Cuddeback said. “That is an important concept in the Million Hearts program as it was in our hypertension collaborative, called Measure Up, Pressure Down.”
The hypertension estimator uses four ranges of age, sex, race and ethnicity, and a number of conditions that are often related to hypertension — diabetes, obesity and chronic kidney disease. The assumption is there may be some patients for whom you are missing the hypertension diagnosis, Cuddeback said. “It is a diagnosis that involves a little judgment. People say, ‘Oh, your blood pressure seems a little elevated. Maybe it is just white coat hypertension.’ They often don’t follow up on the process. That is the concern.”
One of the questions they sought to address is how well the co-morbid conditions are picked up in the claims and EHR data. Although diabetes is diagnosed reasonably well, obesity is not, Cuddeback said. Of 8.9 million patients who qualified for this study, 7 million don’t have any of the co-morbid conditions diagnosed if you look at just one year’s claims data.
“We are recognizing that obesity very seldom appears on a claim compared to how often it is an issue for the patient,” he said. “You would think if we went to the EHR problem list, and did a five-year look-back we would pick up quite a few more, but it turns out that looking at the patient’s problem list doesn’t pick up many more patients.”
It is only when you use the clinical data from the EHR, the BMI score and other clinical data that might reveal diabetes in a patient who doesn't have it diagnosed, that you really begin to get a larger proportion of your patients, he said. Using the five-year look-back at EHR clinical data, more than half of the patients have one, two or three of those chronic conditions. “One of the points we are making is to characterize how many additional patients you get with these co-morbid conditions, depending on which data set you look at,” he said.
I asked Cuddeback if that tells us something about how physicians use the problem list.
“I am afraid it does,” he said. “They are doing pretty well at using the problem list for diabetes, but much less well for obesity. That is one of the big issues about obesity: It doesn’t get a lot of visibility or attention clinically. That is why we are beginning a three-year learning collaborative on obesity. As we surveyed members, they said they don’t really know how to even begin the conversation or what type of program to put together.”
Cuddeback sought to look at how well the participating AMGA members are doing at diagnosing hypertension. “We want to compare the estimated prevalence with the actual prevalence.”
They included patients who didn’t have a diagnosis, but had a stage two blood pressure reading (systolic pressure is 160 mm Hg or higher or diastolic pressure is 100 mm Hg or higher) recorded on at least one day over the past five years or stage one reading (140 to 159 mm Hg or a diastolic pressure ranging from 90 to 99) two days over the past five years.
By including those patients, they got percentages that were well above the highest estimate you would get from the CDC estimator.
Cuddeback and Vaishali Joshi, an AMGA Analytics senior data analyst, are now drilling down on the data from individual medical groups looking for explanations. “We could call these people up and tell them we have an estimator that we applied it to your data, and you don’t seem to be diagnosing hypertension as much as you should be. But that is not terribly useful to them,” Cuddeback said. “We have to be more specific.”
“As we look at individual groups, we want to see if there is a site of care where the pattern may be different in terms of hypertension evidence,” Joshi said. At some of the groups, there is more evidence of stage 2 blood pressure readings but no diagnosis or problem list mention. Maybe the patient only came for one visit or two. What age group are those patients? Are they on Medicare? “That is the kind of drill-down we are working on now,” she said.
There are still specialists who don’t see doing blood pressure screenings as their responsibility. Cuddeback said a physician such as a breast cancer surgeon isn’t expected to manage their patients’ blood pressure, but she is expected to check it on visits and refer them on so it can get managed.
He said AMGA is interested in working further with the CDC team “because they are so interested in taking granular data like this and producing examples of finding patients who are hiding in plain sight. We have the clinical evidence, but they are not getting into a registry because it is not documented in a problem list or a claim. The providers are not getting credit for it in terms of risk adjustment and the patients are not getting the benefit of disease management program the organization has.”
One thing the industry will have to grapple with as it works more with big data is the growth of machine learning, Cuddeback added. Traditionally predictive analytics have used conventional logistic regression modeling and clinicians bought in to the model at least in part because they knew what the coefficients were. They could see how the model worked.
“The problem with machine learning is that those models don’t tell you how they are working inside,” he said. “The meta-issue for healthcare is going to be getting people comfortable with machine learning models that are better at prediction, but we can’t as easily explain how they work.”