The authors developed MetaFemina, an automated evidence-synthesis pipeline designed to evaluate associations between nutrition-related exposures and risk of breast, ovarian, and uterine cancer. The platform combines keyword-based literature retrieval from PubMed with LLM-assisted information extraction and performs random-effects meta-analysis to synthesize pooled effect estimates. The study aimed to build the platform and evaluate its performance against two recent peer-reviewed meta-analyses, and to compare exposure–outcome associations across the three cancer types.
MetaFemina was implemented as a web platform that identifies relevant scientific articles by applying keyword searches to titles and abstracts in PubMed. Extracted articles are processed with large language models to automatically extract study-level information relevant to meta-analysis. Extracted data are synthesized using random-effects meta-analysis. The platform also computes measures commonly used in meta-analysis workflows, including heterogeneity assessment, publication-bias evaluation, and leave-one-out sensitivity analyses.
Additional capabilities reported for the platform include generation of sample size calculations derived from synthesized effect sizes, visual summary outputs, and plain-language interpretations to support users in understanding findings. The platform was developed specifically to address evidence synthesis needs in nutritional epidemiology related to women's cancers.
The authors evaluated MetaFemina by comparing its retrieval of eligible studies to two recently published peer-reviewed meta-analyses that examined folate and vitamin E intake in relation to breast cancer risk. MetaFemina achieved high sensitivity: 81.82% in the folate comparison and 80% in the vitamin E comparison. Beyond reproducing most studies included in those manual meta-analyses, MetaFemina also identified additional relevant articles that had not been captured by the prior manual screening: 27 extra studies in the folate comparison and 13 extra studies in the vitamin E comparison. The report emphasizes the platform's ability to both replicate manual review results and to expand retrieval by finding additional potentially eligible studies.
Using MetaFemina across 226 nutritional exposures, the authors report multiple associations that were significant across cancer types or for pairs of cancers:
These associations are reported as those identified by the automated pipeline; the article presents them as findings emerging from the platform's meta-analyses across exposures and cancer outcomes.
The authors report additional analyses that restricted the evidence base to dietary-intake studies (as opposed to total intake or other measures). After this restriction:
The report indicates that restricting to dietary-intake studies altered the evidence for some exposures, demonstrating the platform's capacity to perform subgroup and sensitivity analyses.
MetaFemina generates several user-facing outputs: pooled effect estimates from random-effects meta-analysis, heterogeneity metrics, publication-bias assessments, leave-one-out sensitivity results, sample size calculations based on synthesized effect sizes, visual summaries, and plain-language interpretations. The platform's workflow uses LLMs to extract study-level information automatically, which the authors argue facilitates more rapid evidence synthesis compared with manual extraction.
The authors conclude that MetaFemina demonstrated high sensitivity in identifying relevant literature, successfully extracts key evidence using large language models, and performs standard, statistically rigorous automated meta-analyses. They propose that the framework may accelerate evidence synthesis in nutritional epidemiology and help researchers with study design, hypothesis generation, and interpretation of emerging evidence linking nutritional exposures to risks of breast, ovarian, and uterine cancers.
The authors state that all data produced in the study are available upon reasonable request to the authors and provide a project URL. Funding for the work included support from the National Science Foundation (NSF DMS 2310955). The authors declared no competing interests. Ethical and author-declaration statements indicate relevant guidelines were followed and that participant consent and trial-registration requirements were addressed where applicable.