` node (the package accepts the OAI-PMH, EFetch ``, Europe PMC and bare `

` shapes), default namespaces are always stripped, and the text is split into the sections where each indicator usually appears: acknowledgments, footnotes / author notes, the body, the methods, the abstract and supplementary material. TXT files are read whole and split into paragraphs. ## Rule-based detection Detection is **rule-based and interpretable**: each indicator is a curated set of regular expressions applied to the relevant sections, rather than a machine learning model. This keeps the output auditable (the matched statement is returned) and reproducible. * **Conflicts of interest.** Detected from structured COI footnotes (`fn-type = "conflict"`), from section titles ("Conflicts of interest", "Competing interests", "Declaration of interest", "Duality of interest"), and from a set of text patterns covering financial relationships, consulting, fees, board membership, patents and explicit "no competing interests" declarations. Honoraria-to-subjects and reference text are masked to reduce false positives. * **Funding.** Detected from the XML `` element, from funding section titles, and from text patterns such as "supported by", "funded by", "grant from / number", named funders and award types. Acknowledged funding is required to use explicit funding language (a funding verb tied to a funder), so a bare mention of an institution or the word "support" is not enough. No-funding declarations are excluded. * **Protocol registration.** Detected from registry identifiers (ClinicalTrials.gov `NCT`, PROSPERO `CRD`, ISRCTN, ANZCTR `ACTRN`, DRKS, IRCT, UMIN, ChiCTR) and from registration phrasing in the methods or footnotes. * **Novelty and replication.** Detected from claim patterns such as "for the first time", "to our knowledge", "novel finding" (novelty) and "replicate", "independently validated", "confirmatory cohort" (replication), with negation filters ("failed to replicate"). * **Data and code sharing.** Detected by a native detector (`.detect_data_code`) built from public repository facts and curated benchmark statements: field-specific accession schemes (GEO `GSE`, SRA / BioProject `PRJNA`, PDB, ArrayExpress, dbGaP, ProteomeXchange, Dryad / Zenodo / figshare DOIs, ...), repository URLs and names, deposit / availability / data-availability-statement language, and supplement and file-format signals. Crucially it distinguishes **sharing** ("data were deposited in GEO") from **reuse** ("data were downloaded from GEO") and excludes "available on request". Code repositories (GitHub, GitLab, Bitbucket) only count as data when paired with a data noun, so a code-only GitHub link is not mistaken for data sharing. * **AI-use disclosure.** Detected from a "Declaration of generative AI" type section title, and from text that names a generative-AI tool (ChatGPT, GPT-4, Copilot, Gemini, an LLM, ...) in a manuscript-preparation context ("used ChatGPT to improve the readability") or in an explicit negation ("no generative AI was used"). A negative lookahead keeps the tool sense of "large language model" out of the writing-object pattern, and AI used purely as a research method (not for writing) is not counted. Only evaluated for 2023 onward. * **Open-access licensing.** Read from the JATS `` element and its license-reference URL, and classified to a canonical identifier (`CC-BY-4.0`, `CC-BY-NC-ND-4.0`, `CC0-1.0`, ...). A Creative Commons or CC0 license (or an explicit open-access declaration) sets `is_open_access`; a CC0 data-waiver is not mistaken for the article license. This is the reuse ("R") dimension of FAIR and feeds the [`rfair`](https://github.com/choxos/rfair) assessment. * **Reporting-guideline use.** Detected when authors state they followed a reporting guideline (the EQUATOR checklists: CONSORT, PRISMA and its extensions, STROBE, ARRIVE, STARD, TRIPOD, COREQ, SQUIRE, CHEERS, CARE, and the wider reportilo list), returning which one. Detection is precision-first: a guideline counts only in a reporting context; common-word acronyms (ARRIVE, CARE, RECORD, ...) require the upper-case form beside a guideline noun; and animal-welfare ("Care and Use of Laboratory Animals"), clinical-practice and non-adherence mentions are excluded. ## Languages Conflict-of-interest and funding statements are detected not only in English but also in **Spanish, Portuguese, French, German and Italian**, using language-distinctive patterns matched on transliterated (accent-stripped) text. The German conflict-of-interest detection rate, for example, rose from 33% to 97% once these were added. The other indicators are English-only for now. # Usage: PMC XML The package ships an example PMC XML file. We use it below; replace the path with your own file to analyze a different article. ```{r} xml_path <- system.file( "extdata", "PMID32171256-PMC7071725.xml", package = "rtransparency" ) ``` ## All indicators at once `rt_all_pmc` returns all ten indicators in one call, together with the matched statement text, the publication `year` and article metadata. ```{r} all_indicators <- rt_all_pmc(xml_path) dplyr::glimpse( all_indicators[, c("pmid", "year", "is_coi_pred", "is_fund_pred", "is_register_pred", "is_novelty_pred", "is_replication_pred", "is_open_data", "is_open_code", "is_ai_pred", "is_open_access", "is_reporting_pred")] ) ``` `is_ai_pred` is `NA` here because this example article predates 2023; for a 2023 or later article it would be `TRUE` or `FALSE`. ## Individual indicators ```{r} coi <- rt_coi_pmc(xml_path) c(is_coi = coi$is_coi_pred, text = substr(coi$coi_text, 1, 120)) ``` ```{r} fund <- rt_fund_pmc(xml_path) c(is_fund = fund$is_fund_pred, text = substr(fund$fund_text, 1, 120)) ``` ```{r} register <- rt_register_pmc(xml_path) register$is_register_pred ``` ## Data and code sharing `rt_all_pmc` already reports `is_open_data` and `is_open_code`; `rt_data_code_pmc` is the focused view that also returns the matched statements. Detection is native and needs no external packages. ```{r} data_code <- rt_data_code_pmc(xml_path) dplyr::glimpse( data_code[, c("is_open_data", "open_data_statements", "is_open_code", "open_code_statements")] ) ``` `rt_all_pmc` and `rt_data_code_pmc` also return `open_data_links` and `open_code_links`: the repository and accession URLs extracted from the statements, ready to pass to FAIR-assessment tooling such as [`rfair`](https://github.com/choxos/rfair). Article metadata (title, journal, identifiers, dates) is available separately via `rt_meta_pmc`. ```{r} meta <- rt_meta_pmc(xml_path) dplyr::glimpse(meta[, c("pmid", "doi")]) ``` ## AI-use disclosure `rt_ai_pmc` reports the publication `year`, the year-gated prediction `is_ai_pred` (`NA` before 2023) and the matched text. The `ai-disclosure` vignette covers this indicator in depth. ```{r} ai <- rt_ai_pmc(xml_path) c(year = ai$year, is_ai = ai$is_ai_pred) ``` # Usage: TXT files The plain-text detectors take either a file path or the text itself, and return the same indicator columns as the XML detectors. `rt_all_pdf()` scores a PDF in one call (it needs the poppler `pdftotext` utility). The chunks below are illustrative and are not executed when the vignette is built. ```{r, eval = FALSE} pdf_path <- system.file( "extdata", "PMID32171256-PMC7071725.pdf", package = "rtransparency" ) rt_all_pdf(pdf_path) # all ten indicators from a PDF article <- rt_read_pdf(pdf_path) # or extract the text once ... rt_coi(text = article) # ... and run any detector on it rt_all(text = article) rt_all("article.txt") # a text file works the same way rt_all_txt_dir("path/to/articles") # a directory of TXT and PDF files ``` `rt_ai` is the plain-text counterpart of `rt_ai_pmc`. A text file carries no reliable publication date, so `rt_ai` applies no 2023 year gate (`is_ai_pred` is always `TRUE` or `FALSE`, never `NA`) and cannot confine the scan to back-matter sections the way the XML detector does. Restrict it to articles published in 2023 or later, and expect a slightly higher false-positive rate on papers that use AI as a research method. # Processing many articles `rt_all_pmc_dir()` runs all ten indicators over an entire directory (or a vector of file paths) in one call, designed for corpus-scale analysis. ```{r, eval = FALSE} # Sequential, in memory res <- rt_all_pmc_dir("path/to/xml") # Resumable and parallel: results are written to a CSV in chunks, a re-run skips # files already recorded, and a malformed file yields an is_success = FALSE row # instead of aborting the run. future::plan("multisession") res <- rt_all_pmc_dir( "path/to/xml", output = "results.csv", parallel = TRUE ) ``` # Summarizing a corpus With one row per article, `rt_summary()` reports per-indicator prevalence with a Wilson confidence interval and a sensitivity/specificity-corrected (Rogan-Gladen) prevalence; `rt_score()` adds a per-article count of openness practices; and `rt_plot()` draws prevalence bars and yearly trends. The `transparency-summary` vignette covers this in depth. ```{r} data(rt_demo) # a small simulated example shipped with the package rt_summary(rt_demo)[, c("indicator", "percent", "adj_percent")] ``` # Downloading PMC XML `rt_fetch_pmc()` downloads full-text XML for PMCIDs, PubMed IDs or DOIs from NCBI (or Europe PMC with `source = "europepmc"`), reusing files already downloaded and reporting which articles have a full-text body. `rt_convert_ids()` maps between the identifier types. ```{r, eval = FALSE} got <- rt_fetch_pmc(c("PMC7071725", "32171256"), dir = "xml") res <- rt_all_pmc_dir("xml") ``` # Beyond presence of a statement A few functions go further than detecting whether a statement exists: * `rt_authors_pmc()` and `rt_funders_pmc()` read tagged JATS metadata: ORCID coverage and CRediT contribution roles, and funders with their Crossref Funder Registry and ROR identifiers and award numbers. * `rt_trial_ids()` extracts registry identifiers from the registration text, and `rt_registration_timing()` asks ClinicalTrials.gov whether each trial was registered before it started. * `rt_fill_coi_pubmed()` recovers conflict-of-interest statements that PubMed records but the full-text XML lacks. * `rt_check_links()` checks whether the data and code links a statement gives actually resolve. * `rt_ethics_pmc()` and `rt_ethics()` detect ethics approval and informed consent statements. They are experimental: not yet validated against hand labels, so they are kept out of `rt_all_pmc()`. # Validation Every indicator is benchmarked against hand labels; the current numbers, their provenance and the scripts that reproduce them are in `inst/benchmark/` and `data-raw/benchmark/` of the source repository, and the README summarizes them. The accuracy table `rt_accuracy` used by `rt_summary()` records each estimate with its validation counts, so the corrected prevalence carries the uncertainty of the validation. # Naming convention and dependencies Functions that operate on TXT files do not end in `_pmc`; functions that operate on PMC XML end in `_pmc`. Data and code detection is implemented natively and no longer requires the `oddpub` or `tokenizers` packages.