How Predictive Modeling Cuts Hospital Readmissions

This article appears in the April 2012 issue of HealthLeaders magazine.

With the looming threat of reimbursement losses for preventable 30-day readmissions, healthcare organizations nationwide are analyzing care transitions in an effort to achieve better outcomes and keep patients from returning to their facilities unnecessarily. While transition programs show promise in helping hospitals reduce their readmission rates, predictive models are also being used successfully in tandem with these programs. Three early adopters of these models are achieving positive results thanks to tactics and technology that identify at-risk patients from the outset of care and influence treatment approaches and the level of transitional care needed.

Parkland Health & Hospital System, Dallas: Data algorithm and readmission rates
Since December 2009, Parkland Health & Hospital System in Dallas has been using what it calls the e-Model, one of the first electronic readmission predictive models of its kind. The organization's center for clinical innovation began development on the electronic predictive model in 2007 with an eye toward making real-time identifications of heart failure patients at high risk for hospital readmission or death. Since then, it has expanded the program for all-cause readmissions.

"We are automating the patient care transaction. It's akin to what finance institutions did in the 1960s and '70s. The challenge is: How can we turn this into data that can help clinicians make real-time decisions that affect outcomes? We're taking information from various sources, and based on it we can say with high probability that [the clinicians] may want to suggest a different course of treatment," says John Dragovits, executive vice president and CFO at the $1.1 billion-net-revenue Parkland Health & Hospital System.

"I think people mistakenly believe that getting an EMR is the end of the process, but it's just the beginning. [Predictive modeling] is showing us how to use this information as a stepping stone … to provide better information to caregivers," he says.

The Center for Clinical Innovations at Parkland received grants from several sources, including the National Cancer Institute and the National Institutes of Health to support its work on the model, which combines 29 data points extracted from its EMR. The data includes physiologic, laboratory, demographic, and utilization variables that can be pulled from a patient's EMR within 24 hours of hospital admission. The comprehensive algorithm has proven to be accurate at predicting readmission or death, says Ruben Amarasingham, MD, director of Parkland's center for clinical innovation and assistant professor of medicine at UT Southwestern. Preliminary results show 33% reductions in readmission of Medicare heart failure patients and 20% reduction in readmissions for all HF patients.

Parkland's predictive model compiles a daily report of all admitted patients and essentially profiles patients and places them into risk categories. Clinicians and case managers are then notified which patients are at highest risk for complications, so those patients can be treated accordingly from the early stages of care, explains Amarasingham.

"There's a lot of value in doing this [modeling] because we have an enormous amount of clinical need and a fixed amount of resources, and that's true for all hospitals," Amarasingham says. "Clinical resources are finite, and that's a real problem."

Based on Parkland's preliminary success with this algorithmic approach to preventable readmissions, it received a grant from the Commonwealth Fund to expand the model to all conditions and across several hospitals including its 968-licensed-bed Dallas hospital, two UT Southwestern hospitals, and five Texas Health Resources facilities. The goal is to build the first electronic readmission model that can be applied to any patient in any hospital where EMRs are available and reduce readmission rates. However, Parkland is also interested in ensuring clinical resources are being focused in the right areas.

Amarasingham, who started the predictive modeling project with a team of four and now has 15 people working on the project, says the use of predictive modeling has been well-received by many of the clinicians. However, while many see the value in having this data in the system, not everyone was on board at first or keen to follow the advice of the algorithm.

"It's a culture change, and the clinicians see the value in it. As computers play a larger role in medical decision-making and the delivery of care, I think attitudes change," Amarasingham says. "We need to see how this model changes care and what the human-to-computer interface in clinical decision-making will be, because it's becoming increasingly impossible for clinicians to keep track of the level of detail—both clinical and social—that's needed in order to arrive at a risk-level assessment. Eventually, I believe physicians will demand this type of predictive modeling technology. Ultimately, clinicians and all healthcare professionals will want to adopt practices to get to that level, including the predictive model."

Mount Sinai Medical Center, New York City: Admissions data and readmission rates
For nearly two years, Maria Basso Lipani, LCSW, coordinator of the preventable admissions care team at Mount Sinai Medical Center in New York City, and Jill Kalman, MD, director of the cardiomyopathy program, associate professor of medicine at Mount Sinai's Cardiovascular Institute, and the PACT medical director, have been using admission history data to identify and target for intervention patients at high risk for readmission. With funding from the United Hospital Fund and assistance from the Department of Health Evidence and Policy at Mount Sinai, the team was able to validate that hospitalization history alone is a reasonable proxy for more formal multivariable regression models in predicting 30-day readmission risk.

PACT, which consists of both a social work–led transitional program and an NP-led medical clinic, enrolled patients based upon data culled from Mount Sinai's existing EMR. A physician from the IT department creates a daily list that identifies hospitalized patients who had a least one admission within the past 30 days or two admissions within the past six months, says Basso Lipani.

Kalman explains that 1,171-licensed-bed Mount Sinai launched the PACT program to reduce exposure to federal readmission penalties and to improve health outcomes through better transition of care. "As part of our evaluation of PACT, we wanted to ensure that the program is truly reaching those who are most likely to benefit from the intervention," Kalman says.

To do that, Mount Sinai's health evidence and policy team developed a risk prediction model for readmission within 30 days using logistic regression. "The higher the score, the higher the risk of readmission," Kalman adds.

Last summer, the predictive model was applied to patients enrolled in the PACT program to determine how many of them were at high risk for 30-day readmission. "Ninety-five percent of PACT enrollees had a risk score greater than 3, meaning that their readmission rate was between 19% and 29%," Kalman says. "If these results can be substantiated through further study, we believe this could have national implications for identifying high-risk patients in real-time."

Mount Sinai is showing early success with its model, too. The PACT program has decreased its 30-day readmission rate from 30% to 12% and its emergency department visits by 63% (over three-plus months), and it has a 90% primary care show rate at seven to 10 days postdischarge for patients enrolled in the program.

Basso Lipani says the core of the transitional program's success is the engagement of patients and families in a discussion of what is uniquely driving readmissions for them. "We've learned that patients with the highest medical utilization, at highest risk for readmission, and with the most fragmented care can be reached and their readmission risk can be reduced through our intervention," she says. "The predictive model has validated not just our method of identifying patients, but our outcomes, too."

In addition to the core PACT team, the organization also is successfully piloting the use of volunteers to serve as extenders for the social workers and NPs. Program volunteers assist patients in making follow-up appointments and retrieve medicine from the pharmacy, helping patients overcome small hurdles that otherwise can have readmission consequences.

Mount Sinai hopes to integrate the risk score into the EMR and use it in conjunction with the transitional social worker's assessment to develop a tiered approach to intervention. "Patients at low risk for readmission may do best with a single follow-up call postdischarge, while a moderate-risk patient may need several calls. This is one way in which the predictive model could have a direct impact on the allocation of resources," Basso Lipani says.

"We wondered if modeling readmissions was going to require us to use more data and create a complex score, but we're validating that a simple [admission history] approach works, and we believe it can be set up easily, regardless of where [an organization] is located, its size, or the level of IT support," says Kalman.

The predictive model and transitional care program is showing promise, so much so that Mount Sinai has four federal proposals and PACT is a part of each of them. "The institution really sees this [approach] as a positive, and it has a desire to see it incorporated into future designs, be it ACO or a medical home," she adds.

Cincinnati Children's Hospital Medical Center: Proactive care and readmission rates
Given the young, average age of the patients at Cincinnati Children's Hospital Medical Center, the decision to create a predictive model program wasn't primarily directed at reducing readmissions, explains Frederick C. Ryckman, MD, senior vice president for medical operations and professor of surgery at CCHMC. Rather it was directed at changing the hospital's approach to care from reactive to proactive. However, the preventable admission rates were positively influenced, he says.

"[Predictive modeling] leads to better communication, better coordination of care, and better outcomes—it's the key to preventable admissions," he says. "Our hypothesis is that a lot of healthcare is very predictable, and if you're able to predict at-risk situations, you can preempt them by building robust mitigation strategies. You can deliver better care, improve [patient] safety, use your capacity and space more efficiently, and create a better patient experience overall by preventing problems."

Several years ago, the organization decided to make the shift from reactive care to proactive care, and Ryckman explains that data was essential. His colleague, Stephen Muething, MD, the organization's vice president of safety, started analyzing and modeling data in specific areas of the hospitals to see patterns that would identify patients at risk for complications.

"We wanted to understand when an event might occur, so we could plan for how to react when an adverse event actually happened. We could come up with a solution to prevent the situation or be better prepared to handle it," Ryckman says.

After gathering three years' worth of data, a team created a model that looked at inpatient units for general pediatrics based on pediatric early warning assessments and created scores using behavior, cardiovascular, and respiratory results. Scores of 3 or above linked to clear action and bedside exam by nurses or physicians, and scores of 7 or above linked to an automatic medical response team call.

The 523-licensed-bed CCHMC uses pediatric early warning scores in its predictive metric within its EPIC system to look at comorbidity, previous history, and risk, and then couples that information with the clinician's knowledge to assess the patient's risk level and put contingency plans in place should the worst-case scenarios develop.

"After doing this for a few years, we have achieved a systemwide approach to using at-risk predictions, and plans are in place to prevent potential risks from occurring," says Ryckman, who notes that the organization also uses the predictive model to determine the level of care coordination needed at discharge.

What the model showed was that while patients could be admitted with a wide variety of problems, there were common potential problems associated with each scenario; for instance, respiratory disease and pneumonia could be complicated by asthma. It would take the clinical staff's input to assess the likelihood of that taking place, and that information would be put into the data to help steer the computer toward deciding the at-risk level of the patient. To help assess the patients, the clinical team on each floor of the hospital meets three times a day (during shift changes) to assess the severity of patients' illnesses.

Also, the morning assessment looks at capacity and the potential for any patient to need intensive care. The staff also does a safety call that includes all the units and alerts the team to potential problems—for instance, if the pharmacy is low on a particular drug. This information is then paired with the technology, which supports the teams. For instance, during flu season the predictive model assesses the potential for added ER personnel and services, and creates targets for monitoring patient progress and when a patient escalation plan should take effect.

"We decided to use technology in a supportive role for the clinical staff, rather than as the solution. I believe other organizations could even run this exact approach in a hospital that has no EMR and the only thing they had was a legal pad and pencil," says Ryckman.

No additional staff was needed to run the predictive model program or coordinate the floor meetings that take place each day, he says. "Having these huddles isn't a hugely time-consuming process, and what comes out of it produces a good ROI," he adds."

The main goal of the program was to predict when children may be showing signs of progressive deterioration in their clinical condition and flag early on which patients may need escalated care. "By using this approach, we've seen the length of stay decrease in ICU, as has the number of critical care codes outside ICU," Ryckman says. The number of overall codes outside critical care averaged 20 events per 1,000 hospital days, with a single-quarter high reached in 2007 of over 40 events per 1,000 hospitals days. It now hovers near 10. "I'd say sending kids home sooner, with a shorter length of stay and not having complications, has an impact on our revenue stream, but the goal is to deliver better outcomes for better overall value of care. This eliminates preventable problems and takes waste out of our system," Ryckman explains.

Ryckman's peers at Parkland and Mount Sinai would agree. While for the most part predictive models require an organization putting some financial investment toward technology, it's not new technology—rather it's an investment in the EMRs they are required to have anyway. Organizations can take advantage of their data now and create predictive models that decrease their preventable admissions and improve outcomes and maximize capacity.

"The ROI with predictive modeling is difficult to characterize and analyze, but if you're preventing multiple admissions, then you're making beds available for other patients," Kalman says.

This article appears in the April 2012 issue of HealthLeaders magazine.

Reprint HLR0412-8