Zuckerberg San Francisco Normal Hospital (ZSFG) has achieved a big turnaround in readmission charges by combining a predictive threat mannequin in Epic, standardized care pathways, and proactive outreach to at-risk sufferers. Lucas Zier, M.D., M.S., director of cardiovascular high quality and outcomes at ZSFG, spoke with Healthcare Innovation about this initiative in addition to efforts to share AI improvements with different security web well being programs.
Healthcare Innovation: May you speak just a little bit about why ZSFG and different security web hospitals have struggled with readmission charges previously and and a few of the monetary repercussions that has?
Zier: Popping out of the Reasonably priced Care Act, one pay-for-performance metric entails readmissions, and the hospital readmission discount program was instituted. I believe it was very effectively intentioned, however the impact was that well being programs that are likely to serve susceptible and underserved sufferers, by definition, look after sufferers with medical challenges, but in addition lots of adversarial social wants. That tends to result in a affected person inhabitants which is extra vulnerable to readmission. The downturn impact of that’s that well being programs that serve our most susceptible sufferers are likely to get penalized most severely for elevated readmission charges.
San Francisco Normal Hospital, as of 2016, had a few of the worst readmission charges within the state of California when in comparison with different security web hospitals. The repercussions of that have been several-fold. One was that it was imperiling about $1.2 million in funding that we have been utilizing for medical care packages. Additionally, we realized that elevated readmission was reflective of some sub-optimal outcomes that we have been having, notably in sufferers with coronary heart failure, which was was our greatest driver of readmission. We had elevated mortality charges in comparison with different security web well being programs, and we additionally had fairness gaps in care. Specifically, the Black/African-American coronary heart failure inhabitants had worse outcomes in comparison with our common inhabitants.
HCI: Did you develop or customise a predictive mannequin inside Epic?
Zier: Sure, we took a multi-pronged method. We localized Epic’s readmission mannequin to our well being system. However we acknowledged that simply offering an end-user with a threat prediction was not going to be adequate to enhance well being outcomes. We took that predictive output and developed a choice help ecosystem. We mixed that with expertise that Epic offers, the place you possibly can primarily hyperlink collectively these logic-based blocks inside Epic. We recreated the center failure tips and we have been in a position to floor choice help to suppliers on the level of care that was personalised to sufferers.
We might make suggestions about medical care and medicines, and we handle each medical wants and a restricted slice of social wants — notably substance use, was an enormous problem in our coronary heart failure inhabitants. One instance of choice help can be if a affected person screens constructive for for methamphetamine use, then we have been in a position to set off a referral to our dependancy care crew. The predictive mannequin allowed us to risk-stratify who’re our highest threat sufferers, so we might floor that data to suppliers after which immediate them to position prioritized referrals for particular person sufferers.
We had a inhabitants well being administration crew for coronary heart failure, however they did not all the time know who to concentrate on. The advantage of the predictive mannequin was that it ran throughout our total coronary heart failure inhabitants, so our crew might take a look at this dashboard that we developed, and see who’s anticipated to be at excessive threat for readmission. That crew might proactively look after these sufferers in the neighborhood to attempt to stop a readmission, earlier than it occurred, versus reacting to a readmission that already occurred.
HCI: Do a sure share of those sufferers not have a major care relationship? And is that a difficulty, so far as following them in the neighborhood?
Zier: That may be a actually good query. Sure, some sufferers do not have a major care relationship, and that is likely one of the potential cease indicators earlier than you discharge sufferers is to ensure they’ve a referral to major care, as a result of we all know sufferers who get referral to and are seen by major care are likely to do higher after they go away the hospital. However one of many advantages of the predictive mannequin is that if a affected person is anticipated to have a excessive readmission threat, then when a supplier locations a follow-up referral for cardiology, that referral will get prioritized inside our referral queue.
HCI: Was there a timeframe over which you studied the impression of this intervention?
Zier: Mainly, we had recognized the issue way back to 2015. At that time, we have been utilizing LEAN methodology to begin to develop some countermeasures. These preliminary pilots have been efficient, however they weren’t scalable. We went by an Epic implementation in 2019 and that is after I began to understand that possibly there’s a chance to make use of the EHR for large-scale, system-wide efficiency enchancment. Our outcomes are actually from 2019 to 2024. It wasn’t a randomized analysis, however we did what we name an interrupted time-series evaluation. Basically we checked out numerous charges of issues earlier than we applied the instrument, after which after we applied the instrument. The precise issues that we checked out have been our readmission charges and our mortality charges amongst our coronary heart failure sufferers. We additionally regarded particularly at outcomes inside our Black/African American inhabitants, as a result of we knew we had vital fairness gaps in these outcomes.
Total, there was a 6% discount in mortality. We additionally in contrast our outcomes to 5 different peer security web hospitals in California, and we discovered that we had, in comparison with these different well being programs, a big discount in mortality amongst our coronary heart failure sufferers. You may by no means be 100% certain except you might have a randomized trial, however we really feel that we are able to confidently conclude that it was the impact of the instrument.
We additionally in contrast our readmission charges to different security web hospitals throughout the identical time interval. And for a lot of of these hospitals, their readmission charges went up, whereas ours went down. At our peak, our readmission fee was about 34% and at its nadir it was about 19 p.c. So we had an over 13% discount in readmission charges, taking us from mainly the worst, relying on the month that you simply take a look at, to top-of-the-line amongst security web hospitals. And primarily, we have inverted this development with our outcomes amongst Black/African-American sufferers and our common coronary heart failure sufferers. We had vital inequities, notably in readmissions, and by 2022, we had utterly inverted that development. There was no distinction in readmission fee.
It has been effectively documented that generally when well being programs attempt to institute readmission discount packages, that comes on the expense of different outcomes. It is not unusual for well being programs, for instance, to see reductions in readmission with a slight improve in mortality, and we have been actually targeted on avoiding that. That was why mortality was an essential metric for us. We have been in a position to cut back readmissions whereas additionally lowering mortality. In our minds, it isn’t an efficient end result if you happen to obtain a well being system metric, however you set your sufferers in danger.
HCI: Let’s return to what you talked about earlier — this consortial work with the opposite security web hospitals. Is that to unfold this specific intervention at different security web hospitals or is it broader than that?
Zier: The overarching thought is that we consider that such a expertise has vital alternative to enhance well being outcomes in well being programs that aren’t notably well-resourced.
This value about $1 million to develop, however as a result of we’re in a position to hit readmission metrics, we saved near $8 million.
One method to this readmission discount program is simply rent 10 extra folks and ship them out in the neighborhood, however security web well being programs cannot try this. We predict there’s an actual function for AI and machine studying to deal with challenges inside these well being programs. However $1 million will not be cheap. We’re fortunate at San Francisco Normal that we now have a basis that helps to help this work. And we now have a relationship with College of California, San Francisco, which creates some infrastructure to assist us to construct a few of these personalized instruments. However most security web well being programs don’t have that kind of help. So the collaboratory was based with the intent of taking a few of the well being programs, like Parkland which has an identical setup to San Francisco, and Grady does as effectively, the place we are able to function an incubator for the event of those applied sciences.
If we will develop them, we should always attempt to disseminate them and share them with different well being programs. It is nonetheless very a lot in its inception. What we’re specializing in proper now’s a generative AI instrument that nearly capabilities as a digital social employee to attempt to handle affected person social and behavioral wants inside these populations. We hope that we are able to disseminate learnings. We are able to additionally disseminate expertise, and we are able to make it accessible to one another at decrease value.