Screening records for a scoping review of demographic equity in medical AI evaluations, with a focus on Middle Eastern and Arabic-speaking populations (pilot round, PubMed)

Version 2 (13 August 2026). This version corrects the exclusion and secondary subtotals deposited in version 1 — 113 excluded and 16 secondary were recorded, where the correct figures at the time of deposit were 110 and 19 — and incorporates a full-text verification pass completed 6 August 2026 in which three of the sixteen included studies were reclassified as secondary evidence. Revised dispositions: 145 screened, 110 excluded, 22 secondary, 13 included (9 evaluating medical AI on Middle Eastern patient populations, 3 on Arabic-language performance, 1 global comparator). The headline proportion is revised from 4 of 11 to 2 of 9 under a strict stratification criterion. The total number of screened records is unchanged; only their classification. Screening records supporting the pilot round of a JBI-framework scoping review examining whether evaluations of medical artificial intelligence report performance stratified by demographic variables, with particular attention to Middle Eastern and Arabic-speaking populations. Protocol registered at https://osf.io/pnwf7 (24 July 2026), amended 4 August 2026. This deposit contains: 1. Screening log, full record — all 145 records screened at title and abstract, each with its call, the abstract language supporting that call, and the author's verification note. Includes the method and provenance table, the screening rubric, the search strategy with its declared limitations, and the screening concordance analysis. Section 7 carries a dated correction block: the stage-1 tally is retained unchanged, with the current dispositions set above it. 2. Search record, 24 July 2026 — the contemporaneous record of both Boolean searches executed against PubMed via the NCBI E-utilities API, with both strings, both date ranges, both total-match counts, and all retrieved records listed. Its decision columns are intentionally empty; screening was conducted through the two-stage process documented in record 1. Unchanged from version 1. 3. Screening results, confirmed — final eligibility dispositions with reasoning, the two-tier eligibility structure adopted at amendment, PRISMA-ScR counts, and the round-2 search strategy. Supersedes the tier assignments file deposited in version 1. 4. Record list — all 145 screened records identified by PubMed ID and search arm. Because the general search retained the top 50 of 889 matches by relevance ranking, and PubMed relevance ranking is not temporally stable, this list rather than the query is the authoritative record of what was screened. Unchanged from version 1. 5. Data extraction table (new in version 2) — 16 rows by 25 columns covering the 13 included studies and the 3 reclassified, every field read from PMC full text, with verbatim source notes for each disposition. 6. Record of AI assistance (new in version 2) — a complete account of AI-assisted decisions including every identified error and its correction: stage-1 agreement 139 of 141, stage-2 full-text verification 13 of 16 upheld, and four errors identified by the model in its own prior output. 7. Figures (new in version 2) — the PRISMA-ScR flow, the reporting matrix, the equity audit and the screening concordance analysis. Deposited as vector PDF: they preview inline without download, stay sharp at print size, and their text layer is machine-extractable, so the cell values can be read directly rather than only viewed. Search strategy. Two Boolean searches were executed against PubMed via the NCBI E-utilities API on 24 July 2026. A general search requiring one of five exact bias phrases combined with artificial intelligence terms and clinical terms, over 2020-2026, returned 889 matches, of which the top 50 by relevance ranking were screened. A regional supplement combining a broader bias and algorithm block with Middle Eastern and Gulf region terms, over 2015-2026, returned 95 matches, all of which were screened. Screening procedure. Title and abstract screening was conducted in two stages. A large language model (Claude Sonnet 5, Anthropic) applied a pre-specified two-question rubric to all 145 retrieved abstracts, recording each call with the abstract language supporting it. The author then reviewed all 145 records against the same abstracts, confirming, overriding, or resolving every call, and retrieving full text where the abstract was insufficient. Every included study was subsequently read in full text; three were reclassified as a result, and the reclassification is documented in record 5. Limitations. These records are a pilot round: one database, one search date, and a relevance-ranked subset. The general query's reliance on exact bias phrases, combined with relevance-ranked sampling and no publication-type exclusion, returned 49 reviews and commentaries against a single primary study, which is a property of the query rather than of the literature. Proportions derived from these records are bounded by the query and the single database that produced them and are not presented as coverage of the wider literature.