Rheumatoid arthritis and elevated cancer risk. Rheumatoid arthritis (RA) is a chronic autoimmune disease affecting approximately 0.5 to 1% of the global population, causing persistent joint inflammation that also involves other body systems. RA patients are known to have a 40 to 60% higher lung cancer risk than the general population according to multiple observational studies and meta-analyses. Yet the nature of this relationship has remained poorly understood, partly because the same factor - smoking - strongly promotes both RA and lung cancer, making it extremely difficult to disentangle cause from confounding.
The confounding problem with smoking. Smoking is the most important risk factor for lung cancer and also an established environmental trigger for RA development, particularly in people with specific genetic variants. When studying whether RA itself causes lung cancer, or whether both are simply consequences of smoking, traditional observational studies struggle because smoking data is often incomplete in medical records and subject to recall bias. Any analysis that cannot fully account for smoking may give misleading results.
Prior studies lumped lung cancer as one disease. Most previous research treated lung cancer as a single entity when studying its relationship with RA. This is a problem because lung cancer actually encompasses biologically distinct subtypes - adenocarcinoma, squamous cell carcinoma, and small cell carcinoma - that have different cell origins, genetic profiles, and risk factors. If RA specifically promotes one subtype but not others, studies that combine all subtypes will dilute and potentially miss the specific signal.
The three-pronged evidence triangulation approach. This study used an evidence triangulation strategy combining three complementary analytical methods: a large clinical cohort study in Chinese patients using propensity score matching to control confounders, machine learning risk prediction models based on clinical features, and genetic causal inference using Mendelian randomization in an independent European population. By approaching the same question from different angles with independent data, this strategy provides more robust conclusions than any single study alone.
Chinese clinical cohort with propensity score matching. The clinical analysis included 8,867 subjects (4,661 RA patients and 4,111 non-RA controls) from a hospital in Tianjin, China, treated between 2014 and 2024. To ensure fair comparison, propensity score matching was used to create 1,552 matched pairs balanced on sex, age, hypertension, diabetes, and hyperlipidemia. This statistical technique makes the RA and non-RA groups as similar as possible on measured characteristics, so that any difference in cancer rates can more plausibly be attributed to RA rather than baseline differences between patients.
Subtype-specific analysis. Rather than analyzing overall lung cancer risk, the study specifically tracked each histological subtype - adenocarcinoma, squamous cell carcinoma, and small cell carcinoma - separately. All lung cancer diagnoses were based on histopathological examination as the gold standard, with multidisciplinary team review for the small proportion of patients who could not undergo biopsy.
Mendelian randomization for genetic causal inference. Mendelian randomization (MR) uses genetic variants - specific DNA sequences inherited at birth - as natural instrumental variables that proxy for lifetime disease exposure without being confounded by lifestyle factors. The study used large-scale genome-wide association study (GWAS) data from independent European populations: 92,044 individuals for RA, 63,053 for squamous cell lung cancer, and 111,752 for smoking. Critically, it performed multivariable MR (MVMR), which simultaneously adjusts for the genetic predisposition to smoking when estimating RA's causal effect.
Machine learning models for risk prediction. Three machine learning approaches - logistic regression, random forest, and XGBoost - were trained on 22 clinical features from the matched cohort to test whether RA patient lung cancer risk could be predicted at the individual level. SHAP analysis identified which features contributed most to predictions. The dataset was split 80:20 into training and test sets, with tenfold cross-validation and upsampling to address the class imbalance created by rare cancer events.
Not all lung cancer subtypes are equally affected. The most important finding from the clinical analysis was striking subtype specificity. After propensity score matching and multivariable adjustment, RA significantly increased squamous cell carcinoma risk (adjusted odds ratio 2.415, 95% CI 1.104 to 5.283, p=0.027). In contrast, RA showed no significant association with adenocarcinoma (p=0.437) or small cell carcinoma (p=0.564). This means RA roughly doubles the risk of squamous cell lung cancer specifically, while having no detectable effect on the other two major subtypes.
Why the selectivity matters biologically. Adenocarcinoma originates from peripheral airway glandular cells, while squamous cell carcinoma arises from central airway squamous epithelium cells. These cancers have different cellular origins, different sets of driver gene mutations, and different tumor microenvironments. The selective association with squamous cell carcinoma suggests that the mechanisms connecting RA to lung cancer - potentially including chronic airway inflammation, specific autoantibodies, or immunosuppressive treatments - specifically promote the transformation of the squamous epithelium rather than glandular cells.
Risk is higher in older patients with moderate inflammation. Restricted cubic spline analysis identified nonlinear relationships between clinical parameters and squamous cell lung cancer risk. Age became a significant risk factor only after approximately 66 years, with risk rising substantially above 80 years. C-reactive protein (CRP) showed a bell-shaped relationship: squamous cell lung cancer risk peaked at moderate CRP levels of 20 to 30 mg/L, with risk declining at very high CRP levels above 50 mg/L. This suggests that chronic moderate inflammation is more pathologically relevant than acute high inflammation.
Association is stable across patient subgroups. Subgroup analyses stratified by age, sex, ESR levels, and CRP levels all showed consistent associations between RA and squamous cell lung cancer without significant effect modification. This stability suggests the association is a robust biological relationship rather than an artifact of a specific patient group within the cohort.
An unexpected protective signal in basic genetic analysis. When the researchers first performed univariable Mendelian randomization - the standard approach that looks at genetic predisposition to RA alone - they found a surprising result: genetic predisposition to RA appeared to be protectively associated with squamous cell lung cancer (odds ratio 0.985, p=0.016). This completely contradicted the clinical observations. Linkage disequilibrium score regression confirmed there was no significant genetic correlation between RA and squamous cell lung cancer genome-wide (rg = -0.019, p=0.806).
Smoking was masking the true relationship. The apparent contradiction arose because smoking is both a trigger for RA and a major cause of squamous cell lung cancer. When genetic variants associated with higher RA risk are also correlated with genetic variants associated with lower smoking behavior (because people with RA-risk genes may smoke differently), simple MR analysis conflates the two effects. The seemingly protective effect of RA was actually a mirror of the negative correlation between RA genetic risk and smoking behavior in the genetic instruments used.
MVMR resolves the paradox. Multivariable Mendelian randomization simultaneously accounts for the genetic predisposition to both RA and smoking as separate exposures. After adjusting for genetic effects of smoking, the analysis revealed the true causal effect: RA genetically increases squamous cell lung cancer risk (odds ratio 1.02, 95% CI 1.00 to 1.03, p=0.046). This reversal from apparent protection to confirmed risk - achieved by statistically separating smoking from RA - demonstrates precisely why multivariable approaches are necessary when confounders like smoking are genetically correlated with the exposure under study.
Concordance across two independent populations. The clinical cohort (Chinese patients, OR 2.415) and the genetic analysis (European ancestry, OR 1.02) both point in the same direction: RA causally increases squamous cell lung cancer risk. The much larger clinical effect size likely reflects the combined action of biological mechanisms plus unmeasured factors including smoking history, occupational exposures, and medication effects not captured by genetic instruments. The authors explicitly caution that these two evidence sources should be considered complementary, not mutually validating, since they come from different populations and study designs.
Limited individual prediction despite strong population association. Despite the robust population-level association between RA and squamous cell lung cancer (OR 2.415), the machine learning models built on 22 clinical features showed only limited ability to predict which individual RA patients would develop cancer. Test set AUCs were 0.684 for logistic regression, 0.671 for random forest, and 0.572 for XGBoost. Ten-fold cross-validation AUCs averaged only 0.64 to 0.73 with high standard deviations of 0.12 to 0.18.
The gap between population statistics and individual prediction. This finding illustrates a common but important limitation in clinical AI: a statistically significant population-level odds ratio does not automatically translate into useful individual-level prediction. A 2.4-fold increased risk at the population level still means most RA patients will never develop squamous cell lung cancer, and routine clinical laboratory values cannot reliably distinguish which individuals will. Cancer development involves stochastic genomic events and environmental exposures not captured by standard clinical tests.
Most predictive clinical features. SHAP analysis across all three models identified triglycerides, blood glucose, erythrocyte sedimentation rate, HDL cholesterol, and C-reactive protein as the most important predictive features. This finding is biologically coherent: metabolic dysregulation and systemic inflammation - both characteristics of active RA - are also mechanistically linked to cancer development. Yet even these features carry insufficient discriminatory power for individual clinical decision-making.
Why models underperformed. The authors attribute limited model performance to three main factors: the small number of squamous cell lung cancer cases (only 32 cases in the matched cohort), the absence of key variables that cannot be measured from routine records (especially smoking history, detailed disease activity scores, and genetic markers), and the complex gene-environment interactions underlying cancer development that require more comprehensive data to capture.
Seven shared causal genes identified exploratorily. Through single-gene Mendelian randomization for both diseases separately and taking the intersection, the researchers identified seven genes with consistent causal effects on both RA and squamous cell lung cancer. Five genes - including POR and EIF3CL - appear to be shared risk factors for both diseases, while two genes - including ANTXR2 - appear protective for both. The authors are explicit that this analysis uses lenient statistical thresholds and only seven genes, making it hypothesis-generating rather than confirmatory.
Potential shared biological pathways. Functional enrichment analysis of the seven shared genes suggested possible involvement in the PI3K-Akt signaling pathway - a regulator of cell survival, proliferation, and metabolism that is commonly dysregulated in both inflammatory diseases and cancers. However, only two of the seven genes map to this pathway, severely limiting the strength of this inference. This connection is explicitly flagged as a preliminary hypothesis requiring experimental validation, not an established mechanism.
Clinical takeaway: targeted squamous cell lung cancer screening for RA patients. The converging evidence supports implementing targeted squamous cell lung cancer screening for high-risk RA patients - particularly those who are older than 66 years or who have moderate chronic inflammation (CRP 20 to 30 mg/L). Standard lung cancer screening programs focus primarily on smoking history; this study suggests that RA itself should be an additional risk criterion for squamous cell carcinoma screening, regardless of other risk factors.
Population validation limitations to acknowledge. A critical limitation is that the clinical evidence came from Chinese patients while the genetic evidence came from European ancestry populations - and these populations may differ substantially in genetic architecture, LD patterns, allele frequencies, and gene-environment interactions. The authors call for dedicated MR studies in East Asian populations as the most urgent next step before these findings can be applied broadly to diverse clinical populations.
Convergent evidence for a specific causal relationship. This study provides two independent pieces of evidence from different populations and methods: a Chinese clinical cohort demonstrating RA significantly increases squamous cell lung cancer risk (adjusted OR 2.415) but not adenocarcinoma or small cell carcinoma, and European genetic causal analysis using multivariable Mendelian randomization confirming RA as a causal risk factor for squamous cell lung cancer (OR 1.02) after adjusting for smoking. Both lines of evidence point in the same direction despite using completely different methodologies and study populations.
The methodological contribution of multivariable MR. A major contribution of this work is demonstrating how multivariable Mendelian randomization can resolve apparent paradoxes created by correlated confounders. The shift from univariable MR showing apparent protection to MVMR revealing true risk - driven entirely by adjusting for smoking - serves as an important methodological example for other researchers studying diseases where smoking is a correlated confounder of the exposure under investigation.
Limitations that temper conclusions. The most important limitations are the unavailability of smoking data in the retrospective clinical records, which means residual confounding cannot be fully excluded from the clinical analysis, and the cross-population inference challenge - genetic findings in European populations may not directly apply to Chinese or other East Asian populations with different genetic architectures. The clinical prediction models also remain insufficient for individual risk stratification, highlighting the need for more comprehensive risk tools that integrate genetic markers and detailed exposure histories.
Clinical implications moving forward. These findings support treating RA - particularly in older patients above 66 years or those with moderate chronic inflammation - as an independent risk factor warranting targeted squamous cell lung cancer surveillance. Future research should validate these findings prospectively, develop dedicated MR analyses in East Asian populations, and build improved risk prediction models integrating genetic markers, disease activity measures, and environmental exposure data to enable more precise individual screening recommendations.