Using big data to target preventable readmissions

Since fall 2012, more than 14,000 patients admitted to one Texas hospital have had a computer program analyze their medical records to help clinicians predict what type of care would improve their outcomes.

The software used at 213-bed Texas Health Harris Methodist Hospital Hurst-Euless-Bedford scans each patient's electronic health record within 24 hours of admission, looking at multiple data elements such as blood-pressure readings and blood-glucose levels.

“It takes all these pieces of data from the EHR, and it has an algorithm, and tells us which patient is at higher risk for heart failure,” said Dr. Susann Land, chief medical officer of the Bedford, Texas hospital.

Armed with 30-day readmission risk scores for heart failure, the hospital is better able to target intensive follow-up care to those patients who need it most. Interventions include prompt callbacks, a follow-up consultation with a cardiologist, a talk with a social worker and the provision of educational materials. “We only have so many resources,” Land said. “We have to pick and choose which (patients) we manage more intensely.”

Provider organizations like the Bedford hospital, which is part of the Dallas-based Texas Health Resources system, are turning to beefed-up predictive analytics tools to reduce their 30-day readmission rates and avoid Medicare penalties under the Patient Protection and Affordable Care Act. The shift to value-based purchasing in the commercial insurance market and the growth of accountable care organizations are also factors. But so, too, is the drive to improve care for its own sake.

Still, many systems using predictive analytics face tough technological hurdles, and research has shown that the existing models still have relatively weak predictive accuracy and so far have yielded only modest results.

The predictive analytics software used at the Bedford hospital was developed at the Parkland Center for Clinical Innovation, now called PCCI, an independent not-for-profit spun off from the Parkland Health & Hospital System in Dallas. Of the patients analyzed by the PCCI system since its adoption at the hospital, 267 patients have been flagged as high risk and another 139 as moderate risk. The software has helped the hospital cut its 30-day readmission rate for heart failure nearly in half, from 23% to about 12%, Land said.

“Can you imagine analyzing 14,000 patients?” she asked. Without a computer, “It's not doable.”

Vast and diverse information sources are being brought to bear on clinical problem-solving at hospitals and clinics across the country. The PCCI model for heart failure, for example, has 29 variables, including social and behavioral factors such as the census tract of the patient's residence, used to estimate socio-economic status, the number of address changes in the past year, the patient's record for keeping appointments, and whether the patient lives alone.

In contrast, the Framingham Risk Score traditionally used to predict the likelihood of coronary heart disease relies on just a handful of demographic and clinical variables—age, gender, two cholesterol measures, smoking status, blood pressure and the use of blood pressure medications.

The increasingly effective use of predictive analytics has been made possible by steady improvements in natural-language-processing technology, which enables software tools to read and analyze physicians' dictated notes and other free text in EHRs.

“We've got more than 150 million text records (in the Regenstrief database), so finding family history until recently wasn't discoverable,” said Dr. William Tierney, president of the Regenstrief Institute in Indianapolis, which has developed medical informatics and EHR technology. That meant a lot of salient information remained locked away from automated use by predictive analytics systems, to the detriment of optimal patient care. “If I can identify someone as high risk for colon cancer, I might be screening them in a different way than for something else,” Tierney said. “My mother died of colon cancer, so I got screened earlier, and indeed, they did find lesions in me.”

Healthcare data analytics firms are starting to compete on their prowess in natural language processing, Tierney said. “It's still at its infancy but it's much better than when it was in the womb a couple of years ago,” he said.

The PCCI system used at the Bedford hospital is an example of this progress. “One of the things this system does that a lot of the others don't is it uses natural language processing,” Land said. “Our docs use free text. They dictate. They do use the EHR almost exclusively. But unless you have someone going into the chart and abstracting the data, you can miss a lot.”

PCCI founder and CEO Dr. Ruben Amarasingham said that after working on a post-doctoral fellowship in bioinformatics and public health at Johns Hopkins University in the early 2000s, “It dawned on me the possibilities of predictive modeling for use by the healthcare system,” he said. When he arrived to work in Dallas, “I pitched it to Parkland and they took a chance on me,” Amarasingham said.

PCCI built a data-analytics platform, including natural language processing, that “that could sit on top of the electronic health record and pull data from it,” he said. As a result, a nucleus of PCCI predictive analytics software users has formed in Dallas, including Texas Health Resources, which was PCCI's first customer; Children's Medical Center; and Parkland itself.

At Parkland, the public hospital where Amarasingham first tested his system between 2008 and 2010, the first target was 30-day readmissions rates for heart failure. Using the predictive analytics system and a combination of patient follow-up activities it triggered, the system initially was able to cut the rate of Parkland patients readmitted to Parkland or any other Dallas-area hospital from 26.2% to 21.2%, according to published reports. Since then, PCCI has developed risk stratification modules for the two other Medicare readmissions core measures—pneumonia and acute myocardial infarction—as well as several other medical conditions and an all-risk module. “It's allowing us to really target those patients who have a high risk and higher need,” said Marilyn Callies, Parkland's vice president of care management.

All high-risk and some medium-risk Parkland patients receive a follow-up phone call within three days of discharge; a follow-up appointment with a primary-care physician or specialist within seven to 10 days; additional follow-up calls when necessary to make sure the treatment plans are effective; and an additional follow-up call after 30 days to make sure they're following their care plan, said Robert Zubrod, Parkland's director of clinical resource management.

Children's Medical Center in Dallas uses PCCI's predictive analytics tool to improve care for pediatric asthma patients. With 2,000 asthma patients, the medical center has had an asthma disease-management program for a number of years, said Summer Collins, its vice president of population health strategies. Adding PCCI's tool provides Children's with “information we never had access to before, she said. With patients' records being scoured and assessed in real time, “We can look at their utilization history, their social history, how many times they've had a payer change in the past 12 months, all of those things allow us to tailor the approach to that patient,” she added.

In Milwaukee, the 14-hospital Aurora Health Care system is using a predictive analytics tool developed by Humedica, a unit of UnitedHealth Group's Optum. Aurora is a member of the AMGA Collaborative, a data-sharing collective organized by the American Medical Group Association that is partnering with Humedica.

With Humedica's help, Aurora launched two predictive analytics pilots focused not only on readmissions, but also on keeping high-risk heart and chronic obstructive pulmonary disease patients out of the hospital. Aurora identified two groups of patients who were eligible for the pilots and living near its 10 participating outpatient clinics.

The two pilot projects each used six registered nurses as health coaches. Starting in June 2013, the heart failure teams oversaw 129 patients, working with 32 providers. Starting in January, the COPD teams looked after 363 patients, working with 32 providers, two pharmacists, one specialist and one home-care worker. For heart failure, readmission rates dropped to 2% for four months of the 2013 program, compared with 14.4% for the same months in 2012.

“This was a very seamless process and it went very, very well,” said Laura Spurr, Aurora's director of medical group operations. Aurora currently is analyzing return on investment for the pilots, but Spurr said the projects already have proved “a huge success.” Aurora is planning to roll out the program for the two conditions across the entire system over the next 12 months.

Follow Joseph Conn on Twitter: @MHJConn