Reproducibility package: Recalls and adverse event reports of AI-enabled medical devices compared with same-category devices (510(k) cohort study)

Reproducibility package for a retrospective cohort study comparing recalls, adverse event reports, and reports of injury or death between devices on the US FDA AI-enabled medical device list and other Class II 510(k) devices cleared in the same product category and year (2004-2024; recall follow-up to September 30, 2026). Version 3.1 accompanies the manuscript "Early recalls of AI-enabled medical devices reassessed within product categories: a 510(k) cohort study". Contents: analysis code (reproduction of a published early-recall comparison across and within product categories, cohort reconstruction from hashed openFDA exports, Cox models with applicant-clustered and shared-event-clustered variance, cause-specific and device-type analyses, manufacturer analyses, outcome-defined calendar-year placebo analysis and simulation, MAUDE linkage, manuscript build), the derived analytic cohort (58,979 submissions), all model outputs reported in the manuscript and appendix, the source-document audit (56 records), and source manifests with URLs, export dates and SHA-256 hashes. Raw FDA files are public and are not redistributed. See README.md for instructions. Corrections in versions 2.9 and 2.9.1. Version 2.3 described the second source-document review as a blinded human review. Both full reviews were AI-assisted: the first with Anthropic Claude Opus 5.5 during analysis (October 28, 2026) and the second with OpenAI ChatGPT 6.1 (October 30, 2026), blinded to the first review. L.-C.C. adjudicated the six disagreements on October 30, 2026. On October 31, co-author H.H.C., who took no part in data processing or adjudication, independently re-coded 20 of the 56 records (all six disagreements plus 14 randomly selected records) without AI tools; her categories matched the final categories for 20/20 of these records. Version 2.9 also counts applicant size using only earlier clearances, adds shared-event clustering, data-preparation scripts and figure sources. Version 2.9.1 synchronizes the manuscript builder with the submitted manuscript (DOI, contributions, adjudication wording) and adds directory-setup steps to README.md. Changes in version 3.0. Adds the reproduction of a published early-recall comparison (share of first recalls within 1 year of clearance) across all eligible submissions and within product-code by clearance-year strata, with applicant-level bootstrap intervals, and a hazard-ratio ladder under successively finer adjustment for device category (code/analysis/v30_claims.py, results/v30_claims.json; exploratory, added November 1, 2026). The manuscript builder reproduces manuscript version 3.0. Changes in version 3.1. Adds code/analysis/v31_scope.py and results/v31_scope.json (exploratory, added November 1, 2026): composition of AI-listed submissions (hardware vs standalone software product codes; AI-listed submissions without a same-category comparator), decomposition of the early-recall comparison into AI and comparator components, recall class linked from FDA device enforcement reports (export of October 21, 2026; SHA-256 in the source manifest), a first Class I recall model, and a quantitative bias analysis for unlisted AI devices among comparators. The manuscript builder reproduces manuscript version 3.1.