Replicating Readmission Reduction Success in the Safety Net

Zuckerberg San Francisco General Hospital (ZSFG) has achieved a significant turnaround in readmission rates by combining a predictive risk model in Epic, standardized care pathways, and proactive outreach to at-risk patients. Lucas Zier, M.D., M.S., director of cardiovascular quality and outcomes at ZSFG, spoke with Healthcare Innovation about this initiative as well as efforts to share AI innovations with other safety net health systems. 

Healthcare Innovation: Could you talk a little bit about why ZSFG and other safety net hospitals have struggled with readmission rates in the past and and some of the financial repercussions that has?

Zier: Coming out of the Affordable Care Act, one pay-for-performance metric involves readmissions, and the hospital readmission reduction program was instituted. I think it was very well intentioned, but the effect was that health systems that tend to serve vulnerable and underserved patients, by definition, care for patients with medical challenges, but also a lot of adverse social needs. That tends to lead to a patient population which is more prone to readmission. The downturn effect of that is that health systems that serve our most vulnerable patients tend to get penalized most severely for elevated readmission rates. 

San Francisco General Hospital, as of 2016, had some of the worst readmission rates in the state of California when compared to other safety net hospitals. The repercussions of that were several-fold. One was that it was imperiling about $1.2 million in funding that we were using for clinical care programs. Also, we realized that elevated readmission was reflective of some sub-optimal outcomes that we were having, particularly in patients with heart failure, which was was our biggest driver of readmission. We had elevated mortality rates compared to other safety net health systems, and we also had equity gaps in care. In particular, the Black/African-American heart failure population had worse outcomes compared to our general population.

HCI: Did you develop or customize a predictive model within Epic?

Zier: Yes, we took a multi-pronged approach. We localized Epic’s readmission model to our health system. But we recognized that just providing an end-user with a risk prediction was not going to be sufficient to improve health outcomes. We took that predictive output and developed a decision support ecosystem. We combined that with technology that Epic provides, where you can essentially link together these logic-based blocks within Epic. We recreated the heart failure guidelines and we were able to surface decision support to providers at the point of care that was personalized to patients. 

We could make recommendations about medical care and medications, and we address both medical needs and a limited slice of social needs — particularly substance use, was a big challenge in our heart failure population. One example of decision support would be if a patient screens positive for for methamphetamine use, then we were able to trigger a referral to our addiction care team. The predictive model allowed us to risk-stratify who are our highest risk patients, so we could surface that information to providers and then prompt them to place prioritized referrals for individual patients. 

We had a population health management team for heart failure, but they didn't always know who to focus on. The benefit of the predictive model was that it ran across our entire heart failure population, so our team could look at this dashboard that we developed, and see who is anticipated to be at high risk for readmission. That team could proactively care for those patients in the community to try to prevent a readmission, before it happened, as opposed to reacting to a readmission that already occurred.

HCI: Do a certain percentage of these patients not have a primary care relationship? And is that an issue, as far as following them in the community?

Zier: That is a really good question. Yes, some patients don't have a primary care relationship, and that is one of the potential stop signs before you discharge patients is to make sure they have a referral to primary care, because we know patients who get referral to and are seen by primary care tend to do better when they leave the hospital. But one of the benefits of the predictive model is that if a patient is anticipated to have a high readmission risk, then when a provider places a follow-up referral for cardiology, that referral gets prioritized within our referral queue. 

HCI: Was there a timeframe over which you studied the impact of this intervention? 

Zier: Basically, we had identified the problem as far back as 2015. At that point, we were using LEAN methodology to start to develop some countermeasures. Those initial pilots were effective, but they weren't scalable. We went through an Epic implementation in 2019 and that's when I started to realize that maybe there's an opportunity to use the EHR for large-scale, system-wide performance improvement. Our outcomes are really from 2019 to 2024. It wasn’t a randomized evaluation, but we did what we call an interrupted time-series analysis. Essentially we looked at various rates of things before we implemented the tool, and then after we implemented the tool. The specific things that we looked at were our readmission rates and our mortality rates among our heart failure patients. We also looked specifically at outcomes within our  Black/African American population, because we knew we had significant equity gaps in those outcomes.

Overall, there was a 6% reduction in mortality. We also compared our outcomes to five other peer safety net hospitals in California, and we found that we had, compared to those other health systems, a significant reduction in mortality amongst our heart failure patients. You can never be 100% sure unless you have a randomized trial, but we feel that we can confidently conclude that it was the effect of the tool. 

We also compared our readmission rates to other safety net hospitals during the same time period. And for many of those hospitals, their readmission rates went up, while ours went down. At our peak, our readmission rate was about 34% and at its nadir it was about 19 percent. So we had an over 13% reduction in readmission rates, taking us from basically the worst, depending on the month that you look at,  to one of the best among safety net hospitals. And essentially, we've inverted this trend with our outcomes among Black/African-American patients and our general heart failure patients. We had significant inequities, particularly in readmissions, and by 2022, we had completely inverted that trend. There was no difference in readmission rate. 

It has been well documented that sometimes when health systems try to institute readmission reduction programs, that comes at the expense of other outcomes. It's not uncommon for health systems, for example, to see reductions in readmission with a slight increase in mortality, and we were really focused on avoiding that. That was why mortality was an important metric for us. We were able to reduce readmissions while also reducing mortality. In our minds, it's not an effective outcome if you achieve a health system metric, but you put your patients at risk.

HCI: Let’s go back to what you mentioned earlier — this consortial work with the other safety net hospitals. Is that to spread this particular intervention at other safety net hospitals or is it broader than that?

Zier: The overarching idea is that we believe that this type of technology has significant opportunity to improve health outcomes in health systems that are not particularly well-resourced. 

This cost about $1 million to develop, but because we're able to hit readmission metrics, we saved close to $8 million.

One approach to this readmission reduction program is just hire 10 more people and send them out in the community, but safety net health systems can't do that. We think there's a real role for AI and machine learning to address challenges within these health systems. But $1 million is not inexpensive. We are lucky at San Francisco General that we have a foundation that helps to support this work. And we have a relationship with University of California, San Francisco, which creates some infrastructure to help us to build some of these customized tools. But most safety net health systems don’t have that type of support. So the collaboratory was founded with the intent of taking some of the health systems, like Parkland which has a similar setup to San Francisco, and Grady does as well, where we can serve as an incubator for the development of these technologies. 

If we're going to  develop them, we should try to disseminate them and share them with other health systems. It's still very much in its inception. What we're focusing on right now is a generative AI tool that almost functions as a virtual social worker to try to address patient social and behavioral needs within these populations. We hope that we can disseminate learnings. We can also disseminate technology, and we can make it available to each other at lower cost.