The presentation was unusual for a conference keynote in that it contained almost no machine learning. A vice president of analytics at a large nonprofit system walked through a three-year effort to build a governed clinical data platform, then spent four minutes on the models it now serves.

The plumbing work was mundane and enormous: reconciling patient identity across four acquired hospitals, backfilling a decade of unstructured pathology reports, establishing a single source of truth for the problem list, and building an access layer that let a data scientist obtain a de-identified cohort in hours instead of months.

"Every model we have shipped since is boring to build," she said. "That is the entire point. If building a model is exciting at your organization, your data is bad."

The reception was warm and slightly rueful. Several attendees noted afterward that this kind of work is nearly impossible to fund on its own, and typically has to be smuggled into the budget under the name of whatever model is currently fashionable.