Public Clinical-Action Pathway Disclosure in FDA-Listed AI-Enabled Medical Devices

This repository contains the public-release analysis code and release-safe data derivatives for the manuscript:

Public clinical-action pathway disclosure in FDA-listed AI-enabled medical devices

Repository URL: https://github.com/wang177777/fda-ai-actionability-public

Transparent Republication Notice

This repository was republished on 29 July 2026 to remove exploratory development materials that were not used in the manuscript analysis. The republication did not change the author-approved submitted manuscript or supplementary information, the final adjudicated labels, the frozen CAP codebook, the submitted tables or figures, or the reported scientific results. The current public tree is the submission-aligned public-release version.

Study Scope

This is a cross-sectional empirical regulatory-science study of FDA-listed AI-enabled medical device submission-level authorizations. The analysis evaluates whether FDA-linked public materials disclose a connected Clinical-Action Pathway (CAP): AI input -> AI output -> intended user/context -> downstream clinical action.

The study does not evaluate confidential FDA review materials, FDA review quality, device safety, device effectiveness, real-world performance, implementation quality, equity, MAUDE outcomes, or recall causality.

Public-Release Contents

• data/public_release/analysis_dataset_public_release_no_quotes_v1.1.csv: submission-aligned analysis-level public-release dataset with long evidence quotes, internal free-text review notes, and obsolete exploratory workflow-status fields removed.
• data/public_release/CAP_coding_public_release_safe_v1.0.csv: public-release-safe CAP coding derivative with raw long quotes removed/truncated according to the project policy.
• results/tables/: main and supplementary table outputs, including the full-sample reviewer-agreement summary reported in Supplementary Table 6.
• results/figures/: main figure PNG/SVG outputs.
• results/source_data/: source data for the main figures.
• codebook/: frozen CAP codebook and variable dictionary.
• scripts/: current submitted-version public-release verification code.
• manuscript/: the author-approved initial-submission manuscript and supplementary-information files uploaded to npj Digital Medicine on 13 July 2026, plus a Markdown rendering of the manuscript text.
• checks/: numerical, citation, and public-release checks.

Coding Workflow and Provenance

As reported in the submitted manuscript, two human reviewers independently assessed all 1,524 submission-level records using FDA-linked public materials and the frozen CAP codebook. Pre-adjudication agreement was calculated from the paired human-reviewer labels; discordant records were adjudicated before dataset lock. Supplementary Table 6 reports the agreement statistics.

The author team subsequently recorded K.Z. (Kaijun Zhang) and W.W. (Weixin Wang) as the two reviewer-confirmers, with the review date recorded as 6 July 2026. X.Y. (Xiaonan Yang), the corresponding/responsible author, certified the final-label confirmation. The author-certified confirmation reconciled all 1,524 submission identifiers and all 10 canonical final-label fields across the submitted manuscript, supplementary materials, and the final GitHub public-release data. Details are provided in docs/final_label_confirmation_provenance.md.

The current default file tree excludes development-only workflow files and draft-generation scripts that are not part of the submitted workflow. The submitted manuscript, supplementary materials, final public-release data, and author-certified final-label responsibility record define the current reported version.

The public-release analysis dataset contains final adjudicated labels. The named confirmation above concerns those final labels and does not itself replace or reconstruct the paired pre-adjudication record-level labels used for the manuscript's inter-rater statistics. Record-level paired pre-adjudication reviewer labels and signed reviewer confirmations are not included in the current public release; aggregate agreement statistics are provided in results/tables/Supplementary_Table_6_reviewer_agreement.csv. Consequently, this public release reproduces the final descriptive outputs but does not independently reproduce the submitted pre-adjudication inter-rater calculation from row-level source labels.

See docs/current_release_scope.md for the authoritative-version hierarchy and public reproducibility boundary.

Excluded Materials

The public repository intentionally excludes raw FDA PDFs, raw FDA HTML pages, raw extracted full text, raw PubMed/ClinicalTrials.gov API responses, cache files, long evidence quotes, manufacturer websites, Google results, MAUDE records, recall records, press releases, news, and marketing materials.

Reproducibility

The main numerical outputs can be checked from the public-release dataset and source-data files. Install dependencies with:

python -m pip install -r requirements.txt

Run python scripts/verify_submitted_public_release.py to verify the submitted-file hashes, the 1,524-record denominator, the 10 canonical final-label fields, the principal CAP distribution, and the released table/source-data values. Internal source-cache and development-workflow scripts are not part of the current public tree. The author-approved manuscript, exact submitted tables and figures, Supplementary Table 6, public-release source data, and final-label confirmation provenance note are included for audit of the reported descriptive outputs and responsibility chain.

Citation

Please cite the associated manuscript when available. A DOI or final citation will be added after publication.

License

Code is released under the MIT License. Public-release tabular derivatives and documentation are released under CC BY 4.0 unless otherwise noted. Raw FDA/public-source documents are not redistributed in this repository.