Skip to content Skip to navigation

More Accurate Insights Into What Drives Readmissions

September 17, 2014
| Reprints
Carolinas HealthCare chief innovation officer talks about bringing predictive analytics to the point of care

My colleague Mark Hagland recently conducted an in-depth interview with two informatics executives at the 42-hospital Carolinas HealthCare System about their efforts to leverage analytics capabilities to support population health initiatives. In that interview, Allen Naidoo, Ph.D., vice president for advanced analytics, mentioned efforts to predict the risk of readmission while the patient is still in the hospital and develop interventions.

To add a little more granular detail about that aspect of their work at Carolinas, here are some notes I took from a nice presentation last week by Jean Wright, M.D., M.B.A., Carolinas’ chief innovation officer. She described how they built the model working with Predixion Software tools and some of the impacts it has already had.

Wright mentioned that Carolinas had already done some work focused on cutting readmissions related to chronic obstructive pulmonary disease with the help of an ONC Beacon grant starting in 2010. That focused initiative lowered COPD readmission rates from 21 percent to 13.5 percent, she said. But that just made Wright curious to tease apart the data and see which factors were having the biggest impact. “I thought it would be helpful to find out which drivers are most impactful on readmissions,” she said.

After defining some critieria, the health system chose to work with Predixion to help build a strong, reliable predictive model of readmissions that could use Carolinas’ own data. One of Wright’s goals was to do the work in a live environment. “Most of my past as an academic physician was with data scientists in dimly lit rooms off campus churning through hordes and hordes of data,” she said. “Was there some way we could change that paradigm and bring that information to where the case manager is making the decisions? It is too little, too late if the person is already sent home, and we realize that he is a high-risk patient.” Most tools that used predictive modeling were things only data scientists used, she said. What about something that had such a friendly user interface that nurses could use it on a daily basis? She also wanted clinical users to be part of the design process. “We believe that people on the sharp end of the spear, closest to the patient, often have the best insights as to what makes a difference.”

Development teams include data scientists, case managers, and those on the front line of clinical care. “This becomes our learning lab,” Wright said. “We are learning about readmissions and how to do point-of-care decision making based on analytics.”

In building that team, she said, Carolinas was committed to a multidisciplinary approach. “We knew nursing was a critical element. I am not a nurse. I am an anesthesiologist and intensivist. But because I have practiced in a team environment, I know how important nurses are to transitions of care, discharge planning and case management. So they were a critical element in planning, design and rollout.”

To build the Patient Centered Point of Care Clinical Decision Support (PCPCS) tool, Carolinas started with 200 clinical, demographic, and procedure-related variables, and then refined that down to the 40 most important variables.

 The rollout began in the summer of 2013. Using the tool, case managers now see patient lists segmented into four readmission risk groups: very high, high, medium and low. Looking at an individual patient’s risk indicators, the clinician can choose from recommended interventions.

The solution is now fully deployed at 13 hospitals. That means that more than 100,000 patients at time of discharge have been managed for 30-day risk of readmission, and more than 123,000 interventions have been applied. More than 200 case managers are using the tools.

Carolinas’ studies show that about 80 percent of the time its model is correct. When they say someone has a high risk of readmission, that patient is probably going to be readmitted.

“When a patient is readmitted, we go back to the case manager and ask how the model predicted the risk of readmission,” Wright explained. “What interventions were recommended and followed with the patient? We then do a root cause analysis on that readmission. Did we score them wrong, not give the right intervention, or did they not follow instructions? Many of us have found it is too easy to say the patient is not compliant. The reality is that is not usually the reason. This allows us to do a deeper dive and get more accurate insights into what is driving that readmission.”

Wright said there have been some beneficial unintended consequences from using the tool. Nurses and case managers now have a better way to level the load across a hospital. Instead of basing rounds on room number, time of discharge or other information, a reliable method of working the list of patients can be developed. Also, case managers can now be deployed based upon the complexity of the patients and their likelihood of readmission.

She sees more promise for PCPCS as a self-diagnostic tool. “In the areas where we have embedded case managers, is it making a difference over areas where they are not available?” she asked. “Over time, we want to figure out, do these interventions really move the needle?”