Patient Report Triage — Multi-Agent System

A LangGraph-based multi-agent pipeline that ingests patient report PDFs, classifies ailments by specialty and severity, routes them to specialist agents in priority order, and loops unresolved cases back to intake for reassessment (with a safety cap that escalates to human review instead of looping forever). Outputs one recommendation PDF per input report.

This is a decision-support prototype, not a diagnostic device. Any real deployment would need clinical validation, human sign-off on every plan, and regulatory review before touching real patient care.

Architecture

                    ┌─────────────┐
                    │   intake    │  Agent 1 (Delegator)
                    │ classify +  │  - parses report text
                    │ build queue │  - extracts ailments, specialty, severity
                    └──────┬──────┘
                           │ (queue sorted severe → major → minor)
                           ▼
                    ┌─────────────┐
              ┌────▶│  pop_next   │
              │     └──────┬──────┘
              │            ▼
              │     ┌─────────────┐
              │     │ specialist  │  Agent 2..N (one per specialty)
              │     │  consult    │  - produces treatment plan, OR
              │     └──────┬──────┘  - flags "can't determine"
              │            │
              │   resolved/escalated   unresolved (retries left)
              │            │                  │
              │            ▼                  ▼
              │     queue empty?        ┌─────────────┐
              │      /        \         │  reassess   │  back to Agent 1
              │   yes          no       │ (re-classify│  with specialist's
              │    │            │       │ w/ feedback)│  feedback
              │    ▼            └───────┴──────┬──────┘
              │ ┌─────────┐                     │
              └─┤ compose │◀────────────────────┘ (pushed back into queue)
                └────┬────┘
                     ▼
                    END → PDF written

The reassessment loop is a genuine cycle in the graph, capped at MAX_REASSESSMENT_ATTEMPTS (default 3) per case — after that, the case is escalated to "requires human physician review" instead of looping forever.

Multiple ailments from one report are processed in severity-priority order (severe → major → minor).

Setup

Install the package (this registers the p-tri and p-tri-ui commands on your PATH):

pip install patient-triage

Or, if you've cloned this project instead:

pip install .            # from inside this project folder
# or, for local development with live-reload on code changes:
pip install -e .

p-tri is exactly python main.py from earlier — same CLI, same flags — just installed as a proper command instead of a script you invoke by path.

LLM backend (swappable — pick one via --backend)

• lmstudio (default): point at a local model served by LM Studio's built-in OpenAI-compatible server (Settings → Developer → Start Server, default http://localhost:1234/v1). Free, runs entirely locally. Set LM_STUDIO_MODEL env var to match whatever model you've loaded in LM Studio.
• anthropic: uses the Claude API. Requires ANTHROPIC_API_KEY env var.
• mock: deterministic canned responses, no model required — useful for testing the graph wiring offline.

Web UI

For a visual alternative to the CLI, p-tri-ui runs a small local Flask server where you can upload reports, trigger processing, and view any PDF — input report or generated recommendation — inline in the browser (using the browser's native PDF viewer, no extra JS library required).

p-tri-ui                                    # http://127.0.0.1:5000
TRIAGE_LLM_BACKEND=anthropic PORT=8080 p-tri-ui   # override backend / port

What it does:

• Upload — choose one or more PDF files, or select an entire folder (via the "Or choose a whole folder" option), and upload them all in one go. Non-PDF files in a folder selection are silently skipped.
• Process — click "Process" next to any un-processed report to run it through the same graph the CLI uses (shared code path — see pipeline.py), or click Process All to run every un-processed report in one click.
• View — click any input report or generated recommendation to load it in the right-hand pane, titled with its actual name (not "(anonymous)").

This is a local, single-user development tool — the dev server it runs on isn't hardened for multi-user or internet-facing use. If you want to expose it beyond your own machine, put a production WSGI server (gunicorn/waitress) and proper authentication in front of it first.

CLI Usage

# Put patient report PDFs in input_reports/, then:
p-tri --backend lmstudio
p-tri --backend anthropic --model claude-sonnet-4-6
p-tri --backend mock              # offline test, no LLM needed

# Custom folders:
p-tri --input-dir my_reports --output-dir my_recommendations

Each <name>.pdf in the input folder produces <name>_recommendation.pdf in the output folder, containing:

• Resolved specialist treatment plans (with clinical reasoning)
• Any cases escalated to human physician review, and why
• A full audit trail of every classification / reassessment step, for a physician to sanity-check the AI's reasoning

A SQLite log (triage_cases.db) records a summary of every run for later auditing.

Project layout

pyproject.toml                    packaging metadata + the `p-tri`/`p-tri-ui` entry points
src/patient_triage/
    config.py                     specialties, severity levels, retry limits, backend config
    schemas.py                    Pydantic/TypedDict data contracts between agents
    llm_backends.py               swappable LLM backend (anthropic / lmstudio / mock)
    utils.py                      JSON extraction helper for LLM outputs
    pdf_utils.py                  PDF text extraction + recommendation PDF generation
    db.py                         SQLite audit logging
    graph.py                      LangGraph wiring (the cyclic state machine)
    pipeline.py                   shared "process one report" logic (used by CLI + UI)
    main.py                       CLI batch entry point (this is what `p-tri` runs)
    agents/delegator.py           Agent 1: classify + reassess
    agents/specialist.py          Agent 2..N: per-specialty consultation
    web/app.py                    Flask web UI (this is what `p-tri-ui` runs)
    web/templates/index.html      upload form, file lists, PDF viewer pane
    web/static/style.css          UI styling
generate_samples.py                dev helper: regenerates the 5 sample reports

Extending

• Scanned/image PDFs: extract_text_from_pdf raises if no text layer is found. Add OCR (pytesseract + pdf2image) as a fallback if your reports come from scanners.
• New specialties: add to SPECIALTIES in config.py — no other code changes needed, since the specialist agent is generic and parameterized by specialty name.
• Persistent service later: graph.py and agents/ are already decoupled from the CLI in main.py, so wrapping build_graph() in a FastAPI endpoint + queue (e.g. Celery/RQ backed by the existing SQLite — or Postgres at that point) is a relatively small step from here.