绘制144种发病危险因素与39种癌症关联图谱：一项AI驱动的系统评价和meta分析

Background: A comprehensive understanding of the associations between multiple risk factors and cancer incidence is crucial for evidence-based cancer control. While many studies have examined specific risk-cancer pairs, none have yet estimated the entire network of risks across cancer types. This study aims to quantity the associations between 144 cancer-related risk factors and the incidence of 39 cancer types. Methods: Using risk records from CanRisk-DB, a well-established repository that employs graph-based retrieval-augmented generation large language model agents within the PICOS-PRISMA framework, we synthesized relative risks (RRs) or hazard ratios (HRs) of cancer incidence from cohort studied between 1980 and 2024. We meta-analyzed the effect sizes using harmonized definitions and both graph-based and inverse variance approaches. The reliability of this artificial intelligence (AI)-driven meta-analysis was validated by comparing our estimated effects with those from published meta-analyses. Results: A total of 2,388 combinations between 144 risk factors and 39 cancer types were identified from CanRisk-DB. Among these, 131 and 120 risk factors were linked to 36 and 33 cancer types in females and males, respectively. Of 144 risk factors, 92.4% were modifiable. 67 factors were identified as causal risk factors only, such as family history of cancer, immunosuppressive agents, non-alcoholic fatty liver disease, and nitrogen dioxide pollution. However, 77 risk factors showed either causal and protective roles across cancer types, such as tobacco use, alcohol consumption, and type 2 diabetes. For instance, alcohol consumption was positively associated with several cancers (e.g., liver [RR = 1.46; 95% CI, 1.27-1.69], breast [RR = 1.09; 95% CI, 1.06-1.12], and colorectal [RR = 1.08; 95% CI, 1.02-1.13]) but inversely with kidney cancer (RR = 0.81; 95% CI, 0.76-0.87). The cancers with the greatest number of associated risk factors were lung (73 factors), colorectal (57), and liver (53). Overall, the majority of cancer types were associated with multiple modifiable risk factors: 34 cancers (87.2%) with at least 5 risks, 28 cancers (71.8%) with at least 10 risks, and 20 cancers (51.3%) with at least 15 risks. The effect sizes in our analysis are highly consistent with those reported in published meta-analyses (Spearman's ρ = 0.93). Conclusion: AI-driven systematic reviews and meta-analyses accurately captured the complex network of associations between cancers and risk factors. Mapping these relationships facilitated a better understanding of the attributable risk of cancer, thereby informing strategies for cancer prevention and control.