Discharge Clarity Coach

A patient-scoped healthcare agent that turns complex discharge instructions into clear, multilingual guidance and grounded follow-up answers.

Comment

Inspiration

Hospital discharge instructions are often clinically correct but practically unusable for patients. They are long, jargon-heavy, and difficult to act on, especially for patients with limited health literacy or a preferred language other than English. That creates a real care gap after discharge: confusion about medications, warning signs, follow-up steps, diet, activity limits, and when to seek urgent help.

We built Discharge Clarity Coach to address that gap inside the workflow clinicians already use. Instead of creating another standalone chatbot, we built a patient-scoped healthcare agent that can be consulted from within Prompt Opinion and that uses the selected patient's actual context to produce clearer discharge guidance.

What it does

Discharge Clarity Coach transforms complex discharge context into clear, plain-language, patient-specific instructions and grounded follow-up answers.

The agent:

  • reads the selected patient's FHIR-backed context and uploaded clinical discharge documents
  • uses a separate knowledge base to control tone, structure, reading level, and safety behavior
  • generates discharge instructions in simpler language
  • supports multilingual output, including Spanish
  • answers follow-up clarification questions while staying grounded in the same patient context

This makes it useful for discharge education, medication understanding, and safer transitions of care.

How we built it

We built the project as an in-platform Prompt Opinion specialist agent with:

  • patient scope
  • A2A enabled
  • FHIR context enabled

The architecture separates two different kinds of knowledge:

  • Patient truth comes from the selected patient's synthetic FHIR record and uploaded patient-specific discharge PDFs.
  • Communication behavior comes from a dedicated knowledge-base collection that guides plain-language writing, response structure, Spanish phrasing, and safety rules.

In practice, the flow is:

  1. A clinician selects a patient in Prompt Opinion.
  2. General Chat consults Discharge Clarity Coach through A2A.
  3. The specialist agent reads patient-specific context.
  4. It returns clearer discharge instructions or answers a follow-up question in grounded, understandable language.

We also created synthetic healthcare artifacts to support the demo, including:

  • FHIR patient bundles
  • patient-specific discharge documents
  • a structured discharge communication knowledge base

Challenges we ran into

The hardest part was not just generating fluent text. It was keeping the output:

  • grounded in the selected patient
  • clinically aligned with the discharge documents
  • easy for judges and end users to read
  • safe when information is incomplete
  • reliable when invoked through Prompt Opinion's agent-to-agent flow

A second challenge was making the system clearly healthcare-native rather than a generic summarizer. That meant designing the agent so that FHIR context, patient documents, and the knowledge base each had a distinct role.

What we learned

We learned that healthcare agent quality depends heavily on context separation. If patient facts and communication guidance are mixed together, outputs become less reliable. Separating patient truth from presentation rules made the agent more grounded and more consistent.

We also learned that in this hackathon, usability inside Prompt Opinion matters as much as model behavior. A strong healthcare agent is not just intelligent; it must be discoverable, invokable, and easy to demonstrate inside the platform.

What's next

Next, we would expand Discharge Clarity Coach in three directions:

  • support more languages and literacy profiles
  • add discharge-risk and medication-specific clarification modes
  • integrate stronger clinician review and approval workflows before patient-facing delivery

Our goal is to make discharge communication safer, clearer, and more personalized without adding friction to clinician workflows.

Built With

  • a2a
  • fhir-r4
  • multilingual
  • pdf-knowledge-base
  • prompt
  • prompt-opinion
  • python
  • reportlab
  • synthetic-healthcare-data
Share this project:

Updates