NoteForge

NoteForge is a proof-first clinical note generation demo for structured device-style measurements. Every supported sentence in the generated SOAP note is tied to an evidence packet with raw-row provenance, and unsupported claims are converted into explicit refusals instead of unsupported narrative.

Why This Matters

Clinical documentation workflows are a poor fit for free-form summarization when the source data is sparse or noisy. NoteForge focuses on a narrower but more defensible problem: convert structured measurements into conservative note text only when the underlying evidence is strong enough to justify the claim.

Key Features

• Deterministic note generation from structured CSV measurements
• Evidence packets with stable packet IDs and raw-row provenance
• Refusal behavior for missing or insufficient evidence
• Built-in evaluator that rejects unsupported or mislinked note text
• Streamlit demo with timeline view, evidence drill-down, and note inspection
• Default sample dataset for quick local startup

Demo

Architecture Overview

flowchart LR
    A["CSV measurements"] --> B["Ingest + validation"]
    B --> C["Timeline table + chart"]
    B --> D["Evidence packet builder"]
    D --> E["Rule-based claim engine"]
    E --> F["SOAP note renderer"]
    E --> G["Refusal engine"]
    F --> H["Evaluator"]
    G --> H
    H --> I["Grounded note + metrics"]

Loading

Technical Stack

• Python 3.13
• Streamlit
• Pandas
• Altair
• Pydantic
• Pytest

Evaluation And Example Output

The repo includes six tests covering evidence construction, rule behavior, and evaluator failure modes. The evaluator explicitly checks that:

• every supported sentence links to known evidence packet IDs
• evidence types match the claim they are supporting
• unsupported note text raises an error instead of silently passing

Example output from the bundled sample CSV:

Objective:
Fever present within last 4 hours. [evidence: PACKET_TEMP_001]

Assessment:
Refusal: elevated_hr not supported. Reason: mean HR 81.0 bpm < 100 bpm. [evidence: PACKET_HR_001]
Refusal: hypoxemia_possible not supported. Reason: mean SpO2 missing; n_samples 0 < 3. [evidence: PACKET_SPO2_001]

Limitations

• The demo only handles structured measurement inputs, not raw clinical notes or EHR exports.
• The note generator is intentionally conservative and covers a small rule set.
• This is not a clinical product and should not be used for care delivery.

Quickstart

py -3.13 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt
streamlit run app.py

The app auto-loads data/sample.csv when no upload is provided.

Or use the bundled scripts:

powershell -ExecutionPolicy Bypass -File scripts/bootstrap.ps1
powershell -ExecutionPolicy Bypass -File scripts/run_demo.ps1

Tests

pytest

Project Structure

noteforge/
  app.py
  app/
  data/
  note_forge/
  scripts/
  tests/

Roadmap

• Expand beyond a single vital-sign rule set into richer multi-signal reasoning
• Add schema adapters for device exports beyond the demo CSV format
• Introduce benchmark scenarios for missingness, conflicting evidence, and temporal edge cases