Voice-native smart forms and GP CCMP

Ilya Beda (Beda Software), Dion McMurtrie (CSIRO), and Liam Barnes (CSIRO) show a three-way collaboration on voice-driven form filling for the GP Chronic Condition Management Plan (GP CCMP) — a complex, multi-section FHIR Questionnaire aligned with AU Core and the AUCDI national digital-health direction. Liam authored the GP CCMP questionnaire, Dion built the Code Search service that converts free text to SNOMED and LOINC codes, and Ilya built the voice-native SDC form filler in Beda EMR.

In the live demo, voice navigation matches spoken words to tab and section names via LLM fuzzy matching. Voice filling converts free text into structured SNOMED codes in near real-time with Code Search plugged in as an MCP tool. The talk is honest about the trade-offs: each voice input triggers an LLM call and a terminology lookup, so latency is the main open problem, and requests are parallelized per questionnaire-item group to keep it usable.

Dion explains that Code Search is designed to know when not to return a code rather than hallucinate one — handling everything from easy matches to adversarial cases where no valid code exists. Benchmarks are not yet published, but the team is interested in open evaluation data as this space matures.

The motivation traces back to the Sparked community AI workshop in May 2026: AI scribes are already in Australian clinics, the review loop is too long, and near-real-time UX is the proposed mitigation.

Building your own voice-native SDC pipeline? Formbox from Health Samurai ships the same SDC-native form authoring and QuestionnaireResponse handling the demo uses, so the AI layer can plug into a production-ready FHIR foundation.