MediVo

MediVo — an AI post-discharge follow‑up agent that calls heart failure patients after discharge, analyses the phone conversation, flags red‑alerts, and prioritizes who needs attention first.

Comment

Inspiration

Heart failure patients are especially high-risk right after discharge, but hospitals struggle to do timely follow-up at scale. Manual outreach creates nurse workload, and the resulting notes are often unstructured—making it hard to quickly see who needs urgent help.

We wanted something clinically relevant and operationally realistic: not an “AI doctor,” but a workflow layer that helps care teams reach vulnerable patients earlier, capture consistent information, and prioritize callbacks.

That led us to MediVo: a focused AI follow-up agent for post-discharge heart failure patients.

What it does

MediVo turns post-discharge follow-up into an end-to-end workflow:

  • Ingests discharge lists (CSV/JSON; demo uses a mock dataset)
  • Builds a prioritized worklist based on time since discharge and protocol priority
  • Runs follow-up calls (simulated for demo; optional live voice via telephony integrations)
  • Uses a protocol-based script (daily weights, meds adherence/understanding, red-flag symptoms)
  • Extracts structured findings (symptoms, med access, med understanding, follow-up status, red flags)
  • Classifies outcomes with guardrails: OK, NEEDS_NURSE, ESCALATE, or NO_ANSWER
  • Generates clinician-friendly summaries and updates the callback queue
  • Exports EHR-ready outputs (FHIR-style JSON and CSV)

Clinical safety constraint: MediVo does not diagnose or provide treatment recommendations—it asks, captures, summarizes, and routes to humans.

How we built it

  • Frontend: React + TypeScript (Vite), Tailwind + shadcn/ui, React Query

    • Dashboard views for queue status, call flow, call summaries, and exports
    • A single configurable API base URL via VITE_API_URL

  • Backend: FastAPI (Python) with a workflow-first API

    • Patient import + worklist endpoints
    • Call endpoints for simulated calls and live call initiation
    • A post-call pipeline: transcript → structured extraction → guardrailed classification → next-step scheduling
    • Optional LLM support (Gemini/OpenAI) for extraction and narrative summaries (feature-flagged)

  • Voice (optional live mode): outbound calling + streaming audio into a conversational agent stack (telephony + conversational voice agent)

  • Monetization (Paid track): we instrumented the workflow with Paid signals (e.g., patient import, follow-up completion, clinical analysis completion, red-flag detection) so usage is tied to a product/order attribution loop.

Challenges we ran into

  • Scope creep

    • Challenge: Our initial idea spanned too many patient groups and workflows.
    • Fix: We narrowed to post-discharge heart failure follow-up, which made the MVP easier to build and easier to judge.

  • Safety vs usefulness in healthcare

    • Challenge: We wanted real workflow value without unsafe “medical advice” behavior.
    • Fix: We constrained the agent to protocol-based intake + triage, with explicit no-diagnosis/no-treatment boundaries and human escalation paths.

  • Monetization under hackathon time limits

    • Challenge: Paid track required a working end-to-end monetization loop.
    • Fix: We integrated Paid directly into core workflow events so signals are generated by real product usage, not a separate demo.

Accomplishments that we’re proud of

  • Built a focused healthcare operations agent, not a generic chatbot
  • Delivered a demoable, end-to-end workflow: import → prioritize → call → summarize → export
  • Implemented guardrailed escalation categories and red-flag detection
  • Added a real monetization instrumentation loop with Paid
  • Kept clinicians in control with a clear human-in-the-loop design

What we learned

Healthcare AI value often comes from workflow design and safety constraints, not maximum model complexity.

  • Narrow scope improves execution and credibility
  • Clear boundaries increase trust (especially in clinical settings)
  • Data matters only when it becomes actionable next steps
  • Monetization is easiest when it’s integrated into the core user journey, not bolted on later

What’s next for MediVo

  • Make protocols clinician-configurable (question sets + escalation rules)
  • Improve risk-based scheduling and escalation pathways
  • Add multilingual support for patient follow-up
  • Extend to other post-discharge pathways (post-op cardiac, oncology, general medicine)

Long-term, MediVo becomes an integration-ready post-discharge follow-up workflow layer that helps care teams scale outreach safely and prioritize the patients who need help most.

Built With

Share this project:

Updates