Background:
Computational medicine indicated that about two-thirds of cancers are attributable to identifiable risk factors such as heredity, environmental exposures, and iatrogenic causes. Despite extensive research on specific risk factors and their links to cancer, the lack of comprehensive evidence synthesis has prevented accurate estimations of preventable cancer cases and the development of comprehensive control strategies. Advances in large language models (LLMs) and artificial intelligence (AI) now enable the systematic integration of all available data to maximize our understanding.
Objectives:
This study aimed to quantity the associations between 134 cancer-related risk factors and the incidence of 39 cancer types.
Methods:
We conducted a systematic review and meta-analysis of cohort studies on cancer risk factors, utilizing harmonized risk records from the AI-driven CanRisk-DB literature database—a repository that employs GRAG-based LLM agents within a PICOS-PRISMA framework. The literature was sourced from PubMed, Embase, and the Cochrane Library, covering publications from 2000 to 2024. To categorize these risk factors, we generated semantic embeddings, applied principal component analysis (PCA) for dimensionality reduction, and performed clustering using the Leiden algorithm. Using standardized definitions, we synthesized relative risks (RRs) and hazard ratios (HRs) via inverse-variance meta-analytic methods. Finally, we evaluated the robustness of our pooled estimates by comparing them with results from published meta-analyses.
Results:
From an initial screen of 435,975 publications, we included 9,550 highly relevant full-text articles, yielding 445,646 structured risk records that span a network of cancer-related risk factors and cancer types across 80 countries. Our meta-analyses identified 1,818 unique associations between 134 risk factors and 39 cancer types. Of these factors, 124 (92.5%) were modifiable. While 51 factors were associated exclusively with increased risk, 73 demonstrated heterogeneous associations across cancer types. For instance, alcohol consumption was positively associated with liver, breast, and colorectal cancers but inversely associated with kidney cancer. Lung, colorectal, and liver cancers had the highest numbers of associated risk factors. The effect estimates from our AI-driven meta-analyses were highly consistent with those from published meta-analyses (Spearman’s ρ = 0.93).
Conclusions:
Using an AI-driven systematic meta-analysis framework, this study presents a comprehensive network mapping of cancer risk factors. The CanRisk-DB-derived network provides a robust foundation for cancer etiology research, population-attributable risk estimation, and the development of evidence-based, targeted prevention strategies globally.

Graphical abstract