From Three States to One Model: AI-Assisted Harmonization of Multi-State Medicaid Promoting Interoperability Data

• In Event: From Text to Evidence: AI-Assisted Analysis and Measurement

Abstract

Comparative policy analysis across states is only as good as the comparability of the underlying data, and Medicaid's Promoting Interoperability program is a case study in just how uncomparable "public" data can be. States were never required to report meaningful use in a consistent format, so a dataset that looks straightforward for one state can be structurally incompatible with another, and can even change shape partway through its own history. This tutorial shares a practical workflow for using an AI assistant to enrich and harmonize Medicaid Promoting Interoperability data from Florida, Georgia, and North Carolina, drawn directly from the presenter's own project. The core example: Florida and Georgia reported a relatively complete row per provider; NPI, year, program year, payment amount, and EHR certification number. North Carolina split its own history in two, recording meaningful use status directly through 2014 and only a payment amount afterward, with no status field at all. Getting an AI assistant to help infer status from payment amount and map program year to meaningful use stage is a concrete, adaptable pattern for anyone whose multi-source data doesn't actually agree with itself. The session also touches on two AI-assisted API scripts used to enrich the data (vendor lookup via the ONC Certified Health IT Product List, provider demographics via NPPES) and one real limitation those scripts couldn't get around, deactivated NPIs that simply aren't retrievable through the API, as an honest example of where AI-assisted enrichment hits a wall. Designed as a fast-moving, example-first Exhibit Hall tutorial rather than a lecture, the session prioritizes showing real prompts and real harmonization logic over discussing AI in the abstract, and leaves real time for attendees to compare notes on their own multi-source data headaches. The goal is not technical mastery but a transferable habit: treating an AI assistant as a collaborator for reconciling inconsistent public data, while staying clear-eyed about what it can and can't fix. This session is aimed at researchers and analysts with some existing AI experience who want practical, adaptable examples from a real project rather than a tools overview and who have their own version of "the data that doesn't match" to bring to the discussion.

Author

• Aaron Rapp, Winston-Salem State University
  Facilitator