Voice AI for PROMs collection

Vitaliy Banchenko shows a different interface for the same SDC contract: instead of a web form, an AI agent calls the patient before the appointment, runs the questionnaire by voice, and returns a QuestionnaireResponse. The talk covers why PROMs collection is worth automating — SMS and email links get missed, contact-center calls do not scale, and filling forms during the visit burns clinician time.

The live demo walks through the pipeline. The agent uses Gemini for the conversation and GPT-5.6 for extraction; configuration bundles the prompt, the set of questionnaires, context variables, and model and voice selection. On extraction the pipeline produces Observations, Provenance, DocumentReference, and Task resources, with Provenance chaining the conversation transcript, the extraction prompt, the extraction device, and every produced resource so any answer can be traced back to the exact spoken words.

Two custom canonical resources — SDCVoiceAgent and SDCVoicePrompt — carry the agent and prompt configuration through versioning. The implementation uses the AI Transparency IG plus additional extensions covering the level of granularity clinical review actually needs.

Voice AI is a Formbox capability, not a side project. Try the voice-native SDC pipeline in Formbox — the same populate-and-extract contract Vitaliy demos, now open to anyone building on FHIR Questionnaire.