nur-dev/camp-er-episode-core-r1a-paper

Audit-Locked Clinical Episode Modeling for Early Sepsis Warning — manuscript and evidence-verification code

This repository contains the code and derived evidence needed to regenerate the manuscript "Audit-Locked Clinical Episode Modeling for Early Sepsis Warning: Deterministic Architecture Adjudication and Held-Out Source-System Evaluation" from the frozen CAMP-ER Episode-Core R1A study.

Every reported number originates from the evidence registry rather than from the manuscript text. Figures and tables are rendering products: they are generated from evidence/data/extracted-evidence.json and are not hand-edited. Changing typography or journal dimensions therefore does not require re-entering any result.

Contents

Path	Contents
main.tex, abstract.tex, sections/	Manuscript source
supplement/	Supplementary manuscript source
bibliography/	Reference database
scripts/extract_evidence.py	Builds the evidence registry from the frozen study artifacts
scripts/generate_assets.py	Generates all figures and tables from the evidence registry
scripts/validate_paper.py	Verifies hashes, claim coverage, placeholders, and forbidden claims
evidence/evidence-manifest.json	Claim-level evidence registry
evidence/data/extracted-evidence.json	Extracted evidence used to render figures and tables
evidence/values.tex	Reported values exposed to LaTeX as macros

Build

Requires Python 3.11+, a TeX distribution providing pdflatex and latexmk (with TikZ, pgfplots, booktabs, natbib, seqsplit), and PyMuPDF for figure rasterization.

make all

This regenerates the figures and tables, validates the package, and compiles main.pdf and supplement/supplement.pdf offline. Each figure is authored as a standalone LaTeX source (TikZ/pgfplots) that is compiled to PDF and rasterized to SVG and PNG.

The evidence target rebuilds the registry from the frozen study artifacts and requires authorized access to those sources, with R1A_PROJECT and R1A_DATASTORE set to the project and training datastore roots. It is not part of the default build; the derived registry is included here so the manuscript can be regenerated without access to restricted data.

Data

Patient-level source data are distributed by PhysioNet under their own access conditions and are not redistributed here. Source-artifact locations in the evidence registry are recorded as stable logical roots; the accompanying SHA-256 digest is the verifiable provenance anchor for each record.

Scope

The selected model is a retrospective research artifact and representation-initialization candidate. It is not production-qualified, not a clinical alert, not a treatment recommender, not a triage system, and not independently clinically validated.