FemHealth.Science Bias Detection Methodology: A Transparency Framework for Sex Equity Analysis in Medical Research Literature

Women have been systematically under-represented in the research that informs their own care. Where they are included, findings are frequently not reported by sex, leaving clinicians and patients unable to tell whether a result applies to women at all. The information needed to judge this is present in the literature, but it is scattered, technical, and absent from every mainstream search tool. This framework defines a reproducible, zero-cost method for surfacing sex-equity signals from biomedical research records at the point of search. For each paper or trial it determines whether sex-disaggregated participation is reported, whether outcomes are analysed by sex, and which populations and reproductive contexts are represented or omitted. It applies a structured "Sex as a Biological Variable" absence check, grounded in NIH policy (NOT-OD-15-102), to make explicit what a study does not report, the gaps that conventional metrics hide. A defining principle is provenance: every signal is labelled with its source (a structured database field, full text, the abstract, or unavailable), and nothing is inferred or fabricated. When a signal cannot be established from the available metadata it is shown as absent rather than guessed, because on a medical product a wrong flag is worse than none. The framework operates across multiple open scholarly databases and is built for transparency and reproducibility: its categories, severity bands, and provenance labels are fully specified, so any surfaced result can be traced to its source. The detailed implementation is documented separately and access-controlled. Validation against a hand-labelled gold set A random sample of 149 abstracts from the live sources was independently labelled by the author for four signals: whether participant sex was reported, whether outcomes were analysed by sex, whether pregnancy was an exclusion criterion, and cohort sex composition. "Sex reported" uses a population-based definition: a study is counted as reporting participant sex when the sex composition of its population is stated, including single-sex (for example female-only) cohorts. Scored against these labels, the pipeline identifies whether a study reports participant sex with a precision of 92.3% and a recall of 78.3% (F1 84.7%). Precision is prioritised: when the tool conveys a study's sex composition it is correct in roughly nine cases out of ten, and it never fabricates a participant count. The interface preserves a distinction the headline figure does not: a positive "sex data reported" signal is shown only when a disaggregated count was recovered, while a study identified as single-sex by design but without a recovered count carries a separate, neutral label. Detection of sex-stratified analysis is an area of active improvement and is the priority for the next iteration. Cohort-composition classification agrees with the labels in roughly two-thirds of cases. These figures come from a small, single-annotator sample and are reported as an honest indication of current performance, not a benchmark claim.