An AI-enhanced evidence-mapping framework for exhaled breathomics in cancer diagnostics: integrating multiple large language models (2005-early 2025).

backgroundExhaled breathomics is a promising non-invasive avenue for cancer detection via volatile organic compounds (VOCs).

methodsThis study presents an AI‑enhanced evidence‑mapping framework that integrates. three large language models (LLMs): Tongyi Qianwen, Hunyuan, and Yi to semantically score relevance (0–10) for studies (2005-early 2025) retrieved by a broad initial query. Deterministic decoding (temperature = 0.0; top_p = 0.1), a unified scoring rubric, and threshold-sensitivity analyses (cutoffs 5–7) were adopted. A stratified random sample (n = 100) underwent double expert labeling with adjudication to validate AI screening.

resultsAfter exclusions, the corpus comprised 2,625 records (2,083 ART; 542 REV). ART were de-duplicated 2,083 to 2,007, then combined with REV (total 2,549) for LLM screening, retaining 808 core publications (31.7%). Against expert labels, the multi-LLM consensus yielded Accuracy 0.92, Precision 0.91, Recall 0.89, F1 0.90, Cohen’s κ 0.82. Bibliometric mapping shows sustained growth, with ~ 40% of top-cited studies centered on lung-cancer VOCs and limited mechanistic interrogation of biomarker origins or stage correlations.

conclusionsThe proposed multi‑LLM consensus framework reliably scale relevance assessment for breathomics literature while maintaining transparency and reproducibility. Findings highlight methodological maturation and collaboration expansion, alongside translational gaps in biological plausibility and cross-platform reproducibility.