Otis (right) and his best friend, Dr. Lucas Zier, director of cardiovascular quality and outcomes at Zuckerberg San Francisco General Hospital
Photo: Dr. Lucas Zier
Before 2017, Zuckerberg San Francisco General Hospital – an urban, academic safety-net hospital within the San Francisco Health Network – struggled with some of the highest 30-day readmission rates in California's safety-net sector.
THE CHALLENGE
The problem was not only clinical but also financial: Missing state and federal readmission reduction targets put $1.2 million annually in at-risk funding in jeopardy, funding critical for clinical care. Equally concerning were pronounced disparities in outcomes: Black/African American patients faced significantly higher readmission rates than the general patient population, underscoring both a quality and equity crisis.
"To understand the scope and root causes of the problem, ZSFG applied Lean methodology to conduct a comprehensive, data-driven analysis," said Dr. Lucas Zier, director of cardiovascular quality and outcomes at ZSFG. "The review revealed that heart failure was responsible for more than 40% of unplanned readmissions, exerting a disproportionate impact on overall performance metrics.
"Modeling suggested reducing HF readmissions could enable the hospital to meet its systemwide readmission targets," he continued. "This insight led to a focused strategy: Concentrate resources on HF, where targeted interventions could be deployed, evaluated and refined more efficiently."
A deeper dive into the drivers of 30-day unplanned HF readmissions revealed three major challenges.
"First, adverse social determinants of health had a powerful influence on outcomes – patients with both HF and methamphetamine use, for instance, were at particularly high risk," Zier explained. "Second, there was no standardized approach to HF care, resulting in wide treatment variation and, at times, care patterns influenced by bias.
"Finally, clinical teams lacked a reliable method to identify which patients were at the greatest risk for readmission, limiting the ability to direct scarce medical and social resources where they were most needed," he added.
PROPOSAL
Building on the findings from this initial analysis, ZSFG launched a six-month pilot on a single inpatient service to test targeted interventions for HF readmission reduction. Two core strategies formed the backbone of the pilot.
First, an evidence-based inpatient checklist standardized care for all HF patients, ensuring each received complete diuresis prior to discharge, social risk-informed medical therapy, and expedited follow-up within seven days in both primary care and cardiology.
Second, staff created a dedicated "Heart Team," bringing together a previously siloed group of HF specialists; primary care providers; and experts in addiction medicine, palliative care and social medicine to deliver coordinated care to the highest-risk patients.
"The pilots demonstrated promising results but also exposed important limitations," Zier noted. "The paper-based checklist existed outside the clinical workflow and the electronic health record, making it cumbersome to use. The Heart Team lacked a systematic method to identify high-risk patients, relying instead on informal referrals that missed many who could benefit from early intervention.
"To address these barriers, we decided to scale the pilots into a hospital-wide program by embedding both interventions into the EHR and creating a centralized digital platform for HF readmission management," he continued. "This would allow seamless integration into provider workflows and enable real-time patient identification through predictive artificial intelligence."
Staff set three design requirements for converting the checklist into a digital tool: It had to be fully embedded in the EHR to avoid workflow disruption; it had to adapt recommendations to each patient's clinical and social risk profile using provider inputs and live EHR data; and it had to automate data collection and processing to streamline decision making and reduce cognitive load.
"To meet these design requirements, we localized an AI model predicting readmission risks to the ZSFG population to provide a framework for readmission risk stratification," Zier explained. "Leveraging existing EHR capabilities, we built a logic-driven, point-of-care decision support interface that delivered patient-specific guideline-based HF treatment recommendations directly to inpatient providers.
"Additionally, we built an HF dashboard within the EHR that displayed predictive AI-derived readmission risk outputs for all patients with HF in real time," he added.
MEETING THE CHALLENGE
Staff developed two strategies for deployment targeting different end users. The first strategy was to target inpatient providers caring for admitted heart failure patients at the point of care. They created a custom "CarePath," a technology within the Epic EHR that allows builders to create complex logic-based algorithms based on tabular EHR inputs and surface clinical decision support.
"We surfaced these patient-specific clinical decision support recommendations through BPAs embedded within a custom-built navigator in the Epic EHR," Zier said. "We additionally alerted providers of a patient at high risk of readmission through this interface by surfacing a BPA, both alerting the clinician of the patient's elevated readmission risk and prompting a prioritized referral to cardiology clinical at discharge.
"In this manner, AI was combined with a logic-based algorithm that recommended specific provider actions to standardize inpatient care," he continued. "In addition, the recommended actions included both guideline-recommended medical care and recommendations about social drivers where appropriate, for example referrals to ZSFG's Addiction Care Team when the underlying algorithm identified active substance use in a patient."
The second strategy was to target the heart failure population health management team, termed the "Heart Team," through a population health dashboard that allowed the heart team to view in real time patients at increased risk of an unplanned readmission within 30 days.
"Prior to the development of this dashboard, the heart team would rely on scattered information from clinical teams about patients recently admitted who were thought to be readmitted," Zier explained. "After the development of this dashboard, the Heart Team used AI predictions to anticipate who would be readmitted, thereby allowing focus on high-risk future events rather than relying on historical events.
"Predictive AI was operationalized using a localized version of Epic's Risk of Unplanned Readmission model, later replaced with an internally developed gradient-boosted tree model incorporating SDOH data," he continued. "Risk scores were surfaced in the decision support interface, prompting providers to initiate high-priority follow-up referrals to cardiology clinic as I just described."
At the population level, a custom HF dashboard displayed risk-stratified patient lists, enabling the Heart Team to proactively manage patients most likely to be readmitted. The system was fully embedded within Epic with no standalone application required.
RESULTS
ZSFG has enjoyed a number of successes with this work. First, reduced readmission rates. All-cause 30-day HF readmission rates fell from 27.9% pre-implementation to 23.9% post-implementation. Compared to peer California safety-net hospitals, ZSFG moved from having the highest readmission rate to the lowest.
Further, it eliminated the equity gap. In 2018, Black/African American HF patients had a 49% higher adjusted odds of readmission compared to other groups. By 2022, this disparity had been fully eliminated, with readmission rates equalized across racial groups.
Then there is improved survival. Post-implementation, all-cause mortality in HF patients decreased by 6%, indicating reductions in readmissions did not come at the expense of patient survival, a common event in readmission reduction initiatives.
And finally, the financial impact. The program allowed ZSFG to meet pay-for-performance readmission targets annually, retaining $7.2 million in at-risk funding over six years on a $1 million development cost – yielding a greater than seven-to-one return on investment.
"It is difficult to determine which parts of this tool were most effective in achieving each outcome," Zier noted. "Ultimately, we believe each component played a role. For example, some patients may have benefited from the standardization of inpatient heart failure care by receiving access to medications and social support that, prior to the deployment of the tool, may not have been offered.
"Other patients may have benefited from the predictive AI component, which led them to receive a prioritized visit in the heart failure clinic after discharge," he continued. "Previously, there was no prioritization, and therefore, high-risk patients had to 'wait in line' for their appointment, which in some cases could be pushed out several weeks."
Early engagement with the health system after discharge undoubtedly led to some of the improvement in outcomes, he added.
"Finally, population surveillance through the population health dashboard, combined with predictive AI, allowed our health system to locate patients in the community at elevated risk of readmission and proactively care for them outside the hospital," he said. "This type of care strategy was simply not possible prior to the deployment of this tool."
ADVICE FOR OTHERS
EHR-integrated predictive models and standardized care pathways can be transformative in addressing readmissions and other high-impact quality metrics, but their success hinges on thoughtful design and implementation, Zier advised.
"First, technology alone will not drive change – tools must be embedded into clinical workflows and paired with clear, actionable steps for end users," he said. "Predictive outputs should be linked directly to provider actions. Simply surfacing risk scores is unlikely to result in meaningful change.
"Second, engagement is critical," he continued. "Early and ongoing collaboration with frontline clinicians ensures tools are clinically relevant, user-friendly and trusted. Incorporating feedback loops and regular orientation sessions helps sustain adoption."
Finally, equity considerations should be built into both model development and workflow design, particularly in safety-net settings where social risk factors strongly influence outcomes, he added. Predictive models that omit SDOH risk embedding bias; conversely, when designed thoughtfully, they can help close persistent care gaps, he said.
"When implemented as part of a system-wide strategy that integrates analytics, workflow standardization and multidisciplinary care, these tools can deliver sustained improvements in quality, equity and financial performance – especially in resource-limited health systems most in need of support," he concluded.
Follow Bill's health IT coverage on LinkedIn: Bill Siwicki
Email him: bsiwicki@himss.org
Healthcare IT News is a HIMSS Media publication.
WATCH NOW: Wisdom on nurses and IT from a health system's nurse CIO
