OBJECTIVES: Non-steroidal anti-inflammatory drugs (NSAIDs) are often used among adults with Osteoarthritis (OA). Research has shown that many cancers are linked to chronic inflammation. NSAIDs may protect against inflammation and can prevent cancer. However, there is paucity of information on effects of NSAIDs on incident cancer risk. We assessed the association of prescription NSAIDs with incident cancer (breast, colorectal, prostate, and non-Hodgkin’s lymphoma(NHL)) among older adults with OA using interpretable machine learning approaches.

METHODS: A retrospective cohort study design with cancer(N=17,824) and non-cancer(N=11,481) patients from Surveillance, Epidemiology, and End Results(SEER)-Medicare linked database and 5% noncancer random sample of Medicare beneficiaries was used. We used the Xtreme Gradient Boosting (XGBoost) classifier with 54 features, to identify leading predictors using feature importance and novel interpretable SHapley Additive exPlanations(SHAP) to assess their associations. Analyses were conducted using Python 3.8.8 with scikit-learn 0.24.2 wrapper interface.

RESULTS: Overall, 30% cancer and 26% noncancer patients used long-term NSAIDs (greater than or equal to 60 days). Mean cumulative NSAIDs days was 110(±120) for cancer and 152(±130) for noncancer patients. Model accuracy(0.84), precision(0.86), recall(0.86), and area-under-the-curve(0.83) metrics were high. Cumulative NSAIDs days was among the top 3 predictors of incident cancer. Zip-level median income, poverty, education, region, age at diagnosis and female sex were among the top 10 predictors. The cumulative NSAIDs days had a complex relationship with incident cancer. Patients without prescription NSAIDs were less likely to be diagnosed with incident cancer. At other levels of NSAIDs use, the relationship was complex and nonlinear.

CONCLUSIONS: The anti-inflammatory effect of NSAIDs on incident cancer is not conclusive and dependent on the cumulative NSAIDs days. Generally, zip-codes with higher levels of poverty and higher percentage of less than high school education were more likely to be diagnosed with incident cancer. These findings highlight the role of social determinants of health on cancer risk.