According to a cardiac quality specialist, AI models combined with standardized care pathways have the potential to significantly improve readmissions and other key quality metrics, but their effectiveness depends on careful design and implementation.
Prior to 2017, Zuckerberg San Francisco General Hospital — an urban academic safety-net facility within the San Francisco Health Network — faced some of the highest 30-day readmission rates among California’s safety-net hospitals.
The Obstacle
The issue at ZSFG was both clinical and financial: failing to meet state and federal readmission reduction targets put \$1.2 million in annual funding—vital for patient care—at risk. Compounding the concern were stark disparities in outcomes: Black/African American patients experienced significantly higher readmission rates than the broader patient population, highlighting a combined quality and equity crisis.
“To fully understand the scope and root causes of the problem, ZSFG employed Lean methodology to conduct a comprehensive, data-driven analysis,” said Dr. Lucas Zier, director of cardiovascular quality and outcomes at ZSFG. “The review showed that heart failure accounted for more than 40% of unplanned readmissions, disproportionately affecting overall performance metrics.”
“Modeling indicated that reducing heart failure readmissions could allow the hospital to achieve its systemwide targets,” he added. “This insight informed a focused strategy: direct resources toward heart failure, where interventions could be implemented, assessed, and optimized more effectively.”
A detailed examination of the factors driving 30-day unplanned heart failure readmissions identified three primary challenges.
“First, adverse social determinants of health strongly impacted outcomes—for example, patients with both heart failure and methamphetamine use faced especially high readmission risk,” Zier explained. “Second, the absence of a standardized approach to heart failure care led to wide variations in treatment and, in some cases, care influenced by bias.
“Finally, clinical teams did not have a dependable way to identify patients at highest risk for readmission, which hindered the effective allocation of limited medical and social resources,” he added.
SOLUTION
Building on insights from the initial analysis, ZSFG initiated a six-month pilot on a single inpatient unit to test targeted interventions aimed at reducing heart failure readmissions. The pilot centered on two key strategies.
First, an evidence-based inpatient checklist standardized care for all HF patients, ensuring complete diuresis before discharge, socially-informed medical therapy, and expedited follow-up within seven days with both primary care and cardiology.
Second, a dedicated “Heart Team” was formed, bringing together previously siloed HF specialists, primary care providers, and experts in addiction medicine, palliative care, and social medicine to provide coordinated care for the highest-risk patients.
“The pilots showed promising results but also revealed key limitations,” Zier noted. “The paper-based checklist was separate from the clinical workflow and the electronic health record, making it cumbersome to use. The Heart Team lacked a systematic approach to identifying high-risk patients, relying on informal referrals that often missed those who could benefit from early intervention.
“To overcome these barriers, we decided to expand the pilots into a hospital-wide program by integrating both interventions into the EHR and creating a centralized digital platform for HF readmission management,” he continued. “This approach allows seamless integration into provider workflows and enables real-time patient identification using predictive AI.”
Staff established three key design criteria for transforming the checklist into a digital tool: it needed to be fully integrated into the EHR to avoid disrupting workflows; it had to tailor recommendations to each patient’s clinical and social risk profile using provider input and live EHR data; and it had to automate data collection and processing to streamline decision-making and reduce cognitive burden.
“To achieve these goals, we adapted an AI model predicting readmission risk specifically for the ZSFG patient population, providing a foundation for risk stratification,” Zier explained. “Using existing EHR capabilities, we developed a logic-driven, point-of-care decision support interface that delivered patient-specific, guideline-based HF treatment recommendations directly to inpatient providers.
“In addition, we created an HF dashboard within the EHR that displayed real-time, AI-derived readmission risk predictions for all HF patients,” he added.
RESPONDING TO THE CHALLENGE
Staff designed two deployment strategies tailored to different end users. The first focused on inpatient providers caring for admitted heart failure patients at the point of care. They developed a custom “CarePath” within the Epic EHR—a technology enabling the creation of complex, logic-based algorithms using tabular EHR data to deliver clinical decision support.
“We delivered patient-specific clinical decision support through BPAs embedded in a custom-built navigator within the Epic EHR,” Zier explained. “Providers were also alerted to high-risk patients via these BPAs, notifying clinicians of elevated readmission risk and prompting prioritized cardiology referrals at discharge.”
“This approach combined AI with logic-driven algorithms to recommend specific provider actions, standardizing inpatient care,” he continued. “Recommendations included both guideline-directed medical management and guidance addressing social determinants—for example, referrals to ZSFG’s Addiction Care Team when the algorithm detected active substance use.”
The second strategy targeted the heart failure population health management team, or the “Heart Team,” via a population health dashboard that enabled real-time identification of patients at increased risk for unplanned 30-day readmissions.
“Before the creation of this dashboard, the Heart Team relied on fragmented information from clinical teams about patients recently admitted and considered at risk for readmission,” Zier explained. “With the dashboard, the Heart Team could use AI predictions to anticipate which patients were likely to be readmitted, allowing them to focus on high-risk future events rather than past occurrences.”
“Predictive AI was implemented using a localized version of Epic’s Risk of Unplanned Readmission model, which was later replaced by an internally developed gradient-boosted tree model that incorporated social determinants of health data,” he continued. “Risk scores were displayed in the decision support interface, prompting providers to initiate high-priority follow-up referrals to cardiology, as previously described.”
At the population level, a custom HF dashboard presented risk-stratified patient lists, enabling the Heart Team to proactively manage those most likely to be readmitted. The system was fully integrated into Epic, requiring no standalone application.
ACHIEVEMENTS
ZSFG has seen several notable successes from this initiative. First, readmission rates dropped significantly: all-cause 30-day HF readmissions fell from 27.9% before implementation to 23.9% afterward. Among California safety-net hospitals, ZSFG went from having the highest to the lowest readmission rate.
The program also closed 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 was fully eliminated, with readmission rates equalized across racial groups.
Survival improved as well. Post-implementation, all-cause mortality among HF patients decreased by 6%, demonstrating that reductions in readmissions did not compromise patient survival—a common concern in readmission reduction efforts.
Finally, the financial impact was substantial. The program enabled ZSFG to consistently meet pay-for-performance readmission targets, retaining \$7.2 million in at-risk funding over six years on a \$1 million development investment—a more than seven-to-one return.
“It is difficult to pinpoint which elements of the tool were most responsible for each outcome,” Zier noted. “Ultimately, we believe every component contributed. For instance, some patients may have benefited from standardized inpatient HF care, gaining access to medications and social support that might not have been provided prior to the tool’s deployment.”
“Other patients likely benefited from the predictive AI component, which enabled prioritized follow-up visits in the heart failure clinic after discharge,” he continued. “Previously, there was no system for prioritization, so high-risk patients often had to ‘wait in line’ for appointments, sometimes for several weeks.”
Early post-discharge engagement with the health system clearly contributed to improved outcomes, he noted.
“Additionally, population-level surveillance via the health dashboard, combined with predictive AI, allowed our team to identify high-risk patients in the community and provide proactive care outside the hospital,” he said. “This type of proactive strategy was not feasible before the implementation of this tool.”
RECOMMENDATIONS FOR OTHERS
Zier emphasized that EHR-integrated predictive models combined with standardized care pathways can significantly reduce readmissions and improve key quality metrics—but only when thoughtfully designed and implemented.
“First, technology alone is not enough to drive change,” he said. “Tools must be embedded into clinical workflows and paired with clear, actionable steps for end users. Predictive outputs should directly guide provider actions, as simply displaying risk scores rarely leads to meaningful improvements.”
“Second, engagement is essential,” he continued. “Early and ongoing collaboration with frontline clinicians ensures that tools are relevant, user-friendly, and trusted. Incorporating feedback loops and regular orientation sessions supports sustained adoption.”
Equity should also be a core consideration in both model development and workflow design, particularly in safety-net settings where social risk factors heavily influence outcomes. “Predictive models that ignore SDOH risk embedding bias,” he noted, “but when designed thoughtfully, they can help close long-standing care gaps.”
“When implemented as part of a system-wide approach that integrates analytics, workflow standardization, and multidisciplinary care, these tools can drive lasting improvements in quality, equity, and financial performance—especially in resource-limited health systems where support is most needed,” he concluded.
