DPH to launch Epic for behavioral health and set AI governance; commissioners press on bias and evaluation

DPH Chief Information Officer and staff presented an information technology update announcing the Epic Behavioral Health Services go-live scheduled for the morning of May 22 and outlined a new governance approach for artificial intelligence in department operations.

Deputy CIO Jeff Scarfia said the Epic behavioral health transition represents three years of planning, with roughly 1,400 users trained and more than 100 technology specialists and additional experienced Epic users staged to provide "at-the-elbow" support during launch. He described operational improvements, including changing the behavioral health access line so callers can leave with an appointment secured on the first call.

CIO Eric Grama outlined a measured approach to adopting AI: DPH will generally "buy" rather than build AI solutions, establish an information governance steering committee and an AI subcommittee, and publish departmental AI policy principles that will guide procurement and evaluation. He emphasized transparency, data provenance and continuous evaluation cycles to detect biases and validate outcomes.

"AI isn't magic," Grama said. "We will make sure that AI systems we make investments in will be explainable and intelligible to a layperson." He cited existing DPH AI uses including cybersecurity event correlation, a RapidAI radiology tool to provide confirmation support for stroke imaging, and predictive modeling work with UCSF to develop a readmission risk score trained on local DPH data.

Commissioners pressed DPH on how it will detect and mitigate adverse bias, who will authorize and ratify AI policy, prioritization of use cases, and cost considerations. Interim SFHN Chief Medical Officer Nida Torrellas said the AI subcommittee's evaluation plan will include stratified outcome metrics (for example, by SOGI data and homelessness status) so the department can monitor whether models produce disparate outcomes for populations of concern.

DPH staff described near-term AI use cases including Epic/Microsoft-assisted drafting of MyChart patient message replies, optical character recognition for insurance card intake and, later, ambient-note transcription to reduce clinician documentation burden. Staff said they expect to return to the commission with final policy documents in the coming weeks.