Discharge Summary Agent
Agentic system that drafts a clinically-safe discharge summary from messy scanned clinical notes, with a no-fabrication guardrail.
- ROLE
- Builder
- PERIOD
- 2026
- STATUS
- Published
- CODE
- GitHub ↗
OVERVIEW
The Discharge Summary Agent reads a patient's scanned clinical notes and produces a structured discharge summary draft for clinician review. Its defining constraint is that it never invents a clinical fact: any value it cannot source from the documents is marked MISSING, PENDING, or CONFLICT and raised as a flag, rather than filled with a plausible guess. The system was built from scratch without an agent framework as a take-home project, and includes a second component (Part 2) that learns from doctor edits to reduce the editing burden on future drafts while keeping safety flags intact.
ARRIVED AS
Scanned clinical records (71 pages of typed printouts and handwriting) have no text layer, contradict each other across pages, and contain data that is missing or pending. A system reading them must surface conflicts and gaps rather than resolve them with plausible guesses.
WHAT I BUILT
- 01From-scratch agent loop (no framework): each step the model receives the running conversation plus tool schemas, emits a reasoning note and tool calls, and re-plans based on results.
- 02Two-stage ingestion: pypdfium2 rasterises each page, then a vision model (Gemini) transcribes it; unreadable text is marked [illegible] and flagged, never guessed.
- 03No-fabrication enforced at three layers: typed state requires a source page and verbatim quote per value; the citation guardrail rejects writes where the quote is not found on that page; an independent verifier LLM re-reads every recorded value after the loop and removes anything unsupported.
- 04Conflicts, pending results, and missing fields are first-class states (CONFLICT, PENDING, MISSING), stored and surfaced in flags.md, never silently resolved.
- 05Medication reconciliation diffs admission vs discharge lists; any change without a documented, quote-backed reason is flagged automatically.
- 06Provider abstraction lets vision OCR (Gemini) and reasoning (Gemini or DeepSeek) run independently, with automatic key rotation on quota errors.
WHAT CHANGED
- On the 71-page patient record the agent ran 66 steps, raised 41 flags, and the citation guardrail rejected 29 attempts to record a value whose quote was not verbatim on the cited page.
- Surfaced a principal-diagnosis conflict (gastroenteritis vs DKA) and a demographics conflict (mismatched pronouns across pages) rather than picking one answer.
- Post-hoc verifier independently checked 39 recorded values; dropped fabricated or unsupported values and replaced them with flags.
- Part 2 learning loop: mean edit burden dropped from 0.154 baseline to ~0.036 (roughly 80% lower) on a held-out split, with safety retention at 100% throughout.