An automated evidence-synthesis platform called MetaFemina, integrating large language model (LLM)-assisted extraction with random-effects meta-analysis, screened 226 nutritional exposures across breast, ovarian, and uterine cancers. Lutein and beta-carotene emerged as significantly associated with lower risk across all three cancer types, while vitamin D, antioxidants, and soy linked to reduced breast and ovarian cancer risk. Calcium and folic acid were inversely associated with breast and uterine cancers. Conversely, iron, red meat, and copper correlated with higher risk in breast and uterine cancers. Notably, omega-6 fatty acids showed divergent signals — higher breast cancer risk but lower ovarian cancer risk — underscoring the complexity of dietary fat research. The platform achieved 80–82% sensitivity versus peer-reviewed manual meta-analyses, while retrieving 13–27 additional studies human screeners missed.
This preprint, not yet peer-reviewed, represents a methodologically ambitious attempt to industrialize nutritional epidemiology synthesis — a field historically bottlenecked by slow, labor-intensive manual review. The carotenoid findings (lutein, beta-carotene) align with existing mechanistic literature on antioxidant-mediated DNA protection and estrogen metabolism modulation, lending biological plausibility. However, all associations remain observational and confounding in nutrition studies is notoriously difficult to eliminate. The platform's LLM extraction layer introduces novel error modes — hallucination, inconsistent unit parsing — not yet fully characterized. Clinically, the calcium, vitamin D, and soy findings warrant cautious optimism but not dietary prescriptions pending validation. The tool's real value may lie in hypothesis generation rather than replacing rigorous human-led meta-analysis.
