Zuckerberg San Francisco General Hospital

As of 2016, ZSFG had some of the worst heart failure readmission rates in California among safety net hospitals, with a peak readmission rate of ~34% and elevated mortality rates. The hospital faced ~$1.2M in at-risk funding tied to readmission performance metrics under the Hospital Readmission Reduction Program. Significant equity gaps existed, with Black/African-American patients experiencing worse outcomes than the general heart failure population.

ZSFG localized Epic's readmission risk model to their population and built a decision support ecosystem on top of it using Epic's logic-block tooling. The system surfaced personalized, guideline-based recommendations to providers at the point of care — including medication guidance and substance use referral triggers. A population health dashboard enabled care teams to proactively identify and outreach high-risk patients in the community before readmissions occurred, and high-risk patients received prioritized cardiology referral queue placement.

From 2019 to 2024, readmission rates dropped from ~34% to ~19% — an over 13-percentage-point reduction — moving ZSFG from among the worst to one of the best-performing safety net hospitals in California. Mortality among heart failure patients fell by 6%, and by 2022 the readmission equity gap between Black/African-American patients and the general population was fully eliminated. The program cost approximately $1M to develop but generated close to $8M in savings.

• Pairing a predictive risk model with embedded clinical decision support — rather than just surfacing a risk score — is critical to driving provider action and measurable outcomes.
• Safety net hospitals can achieve outsized ROI from AI-driven care programs; a $1M investment yielded ~$8M in savings while also reducing mortality and eliminating equity gaps.
• Sharing AI innovations across a collaboratory of safety net hospitals can democratize access to costly tools that individual resource-limited systems could not build independently.