Conversational AI for Disease Management: What Nature's AMIE Study Means for Previsit
— by Vivax
On 17 June 2026, Nature published Google Research and DeepMind's AMIE study on conversational AI for disease management — a dual-agent system that matched 21…
On 17 June 2026, Nature published 'Towards Conversational AI for Disease Management,' a study from Google Research and Google DeepMind that advances their medical AI, AMIE (Articulate Medical Intelligence Explorer), from one-off diagnosis to ongoing disease management. In a randomized, blinded virtual OSCE across 100 multi-visit case scenarios built on UK NICE and BMJ Best Practice guidelines, AMIE was compared with 21 primary care physicians — and matched them on overall management reasoning while scoring significantly higher on the precision of its care plans and their alignment with clinical guidelines.
What makes the result interesting is the architecture. AMIE uses two agents working together: a user-facing Dialogue Agent that handles the conversation — gathering the patient's story, addressing concerns, and building rapport — and a Management Reasoning (Mx) Agent that works deliberately in the background, reasoning over the patient's history across visits together with hundreds of pages of clinical guidelines and drug formularies at once, using Gemini's long-context capabilities. Separating empathetic dialogue from careful, guideline-grounded reasoning is what let AMIE plan investigations, treatments, and prescriptions reliably across multiple visits.
This is exactly the future we are building toward at Vivax. Our Previsit agent is a conversational AI that meets the patient before the consultation — a phone check-in and an in-clinic previsit booth that turn the patient's own words into a structured, voice-signed chief-complaint report that reaches the physician before the patient even enters the room. The Nature study validates our core premise: a well-designed conversational AI can conduct a careful clinical dialogue and produce structured, guideline-aware output that supports the physician rather than replacing them.
The lesson we take from AMIE is the value of separating dialogue from reasoning and grounding every step in trusted clinical knowledge. Previsit already structures intake per specialty, with separate examination-knowledge bases so the questions match the department; the next step on our roadmap is longitudinal management reasoning across visits — the same multi-visit care planning AMIE demonstrates. Both share the same guardrails: cited, guideline-aligned outputs, and privacy by design. In Previsit, the raw patient audio is never stored — only the structured report and the cryptographic signature that proves its integrity.
Nature's authors are careful, and so are we: AMIE was evaluated in a virtual OSCE with trained patient-actors, not yet in live clinics, and real-world translation will require prospective study under real constraints — work Google says it has already begun with clinical partners. We hold ourselves to the same standard. Vivax builds conversational clinical AI that earns trust through grounding, transparency, and rigorous evaluation — and a peer-reviewed result in Nature is strong evidence that the previsit vision is plausible, and that the entire field is now moving in the same direction.