Zuckerberg San Francisco General Hospital retains $7.2M by cutting heart failure readmissions with Epic-integrated predictive AI

Before 2017, ZSFG had some of the highest 30-day readmission rates in California's safety-net sector, putting $1.2M in annual at-risk funding in jeopardy. Heart failure alone accounted for more than 40% of unplanned readmissions. Black/African American patients faced a 49% higher adjusted odds of readmission compared to other groups, signaling both a quality and equity crisis. Clinical teams lacked a reliable method to identify which patients were at greatest readmission risk, and there was no standardized approach to heart failure care.

ZSFG built a fully Epic-integrated CarePath delivering patient-specific, logic-driven clinical decision support at the point of care, including guideline-based HF treatment recommendations and social-determinant-aware referrals. Predictive readmission risk scores—initially using Epic's Risk of Unplanned Readmission model, later replaced by an internally developed gradient-boosted tree model incorporating SDOH data—were surfaced via BPAs to inpatient providers. A real-time population health dashboard allowed the multidisciplinary "Heart Team" to proactively identify and manage the highest-risk patients without any standalone application outside Epic.

All-cause 30-day HF readmission rates fell from 27.9% pre-implementation to 23.9% post-implementation, moving ZSFG from the highest to the lowest readmission rate among peer California safety-net hospitals. The racial equity gap was fully eliminated by 2022. All-cause HF mortality decreased by 6%. The program retained $7.2M in pay-for-performance funding over six years against a $1M development cost, yielding a greater than 7:1 return on investment.

• Predictive outputs must be linked directly to provider actions—surfacing risk scores alone without embedded workflow steps produces little meaningful change.
• Early and ongoing collaboration with frontline clinicians is essential to ensure tools are trusted, user-friendly, and sustained over time.
• SDOH data must be incorporated into model development and workflow design; omitting social risk factors embeds bias, but thoughtful inclusion can close persistent care gaps.