Inspiration
Healthcare's riskiest moment isn't surgery. It's the handoff. A pending lab result, an allergy mentioned out loud, a "watch this patient overnight" note: these get lost in shift changes, ER-to-ward transfers, and hospital-to-home discharges. Most tools store information, but none verify that the next person actually understood it. We wanted to build the safety checklist healthcare never had, the way pilots confirm every critical item before takeoff.
What it does
Care Handoff Guardian sits between the outgoing and incoming clinician and:
- Auto-drafts the handoff from notes, vitals, and pending orders, highlighting what changed in the last 12 hours.
- Detects gaps and contradictions, like an unowned pending lab, a medication that conflicts with a note, or a worsening trend nobody flagged.
- Closes the loop. The receiving clinician must acknowledge each high-risk item, and unresolved "open loops" stay visible until someone owns them.
- Patient-side brief: a plain-language, multilingual discharge summary with a "what to watch for" timeline.
How we built it
- Data: synthetic patient records and generated clinical notes, so no real patient data is involved.
- LLM pipeline: summarizes the chart into a structured handoff.
- Rule-based safety layer: deterministic checks for missing owners, conflicting medications, and unacknowledged critical items, so safety doesn't depend on the LLM alone.
- Frontend: a web app with an acknowledgment workflow and an open-loops dashboard.
- Stack: [ e.g. Python, FastAPI, React, Claude API]
We scored each handoff by how many seeded errors it caught:
$$\text{Catch Rate} = \frac{\text{errors detected}}{\text{errors seeded}}$$
Our result: [X%] versus [Y%] for a plain LLM summary on [N] test cases.
Challenges we ran into
- Trust vs. noise: too many alerts cause alert fatigue, so we had to rank items by risk and surface only what matters.
- Hallucination risk: a summary that invents or omits a detail is dangerous, so every flagged item links back to its source in the chart.
- No real data: privacy meant building realistic synthetic cases and deliberately seeding errors to test against.
What we learned
- Healthcare failures are often communication failures, not knowledge failures.
- Pairing LLMs with deterministic rules is safer than relying on either alone.
- Measuring what a tool catches matters more than how polished the summary looks.
What's next
- Pilot with clinicians for real workflow feedback
- Integrate with EHR standards like FHIR
- Expand multilingual patient briefs, especially for Arabic and other Gulf-region languages
- Formal clinical validation and privacy compliance
:this is a prototype using synthetic data, not a certified medical device.*