Machine Learning Identifies Community-Level Factors Associated with Lung Cancer Incidence

Lung cancer is the leading cause of cancer-related death in the United States. Individual-level risk factors such as tobacco use are well characterized, but the community-level environmental, social, and economic contributions to geographic variation in incidence remain incompletely understood.

We conducted an ecological analysis of 542 Medical Service Study Areas (MSSAs) in California, linking age-adjusted lung cancer incidence rates (AAIR, 2015-2019) to 27 indicators of pollution, population characteristics, and adult smoking prevalence. Machine learning and SHapley Additive exPlanations (SHAP) distinguished high- versus low-incidence MSSAs; ordinary least squares (OLS) regression quantified covariate-AAIR associations.

A Support Vector Machine (radial basis function kernel) trained on 15 autoencoder-selected indicators discriminated best (cross-validation AUC = 0.891), outperforming a smoking-only model (AUC = 0.685). By SHAP, smoking, linguistic isolation, neighborhood SES, unemployment, and housing burden contributed most. In OLS regression, neighborhood SES quintile (standardized β = -0.57) and Hispanic population proportion (β = -0.51) were the strongest predictors of lower AAIR; higher Black population proportion was associated with higher AAIR. Subpopulation-stratified models performed better for White-specific (AUC = 0.926) than Hispanic-specific AAIR (AUC = 0.744).

Community-level smoking, pollution, and population characteristics data improved the identification of regions with elevated lung cancer risk.

Community-level indicators can identify areas of elevated lung cancer burden, including the many where registry rates are suppressed. These associations are concurrent rather than etiologic, and no causal interpretation is intended. Lower predictive performance in some subpopulations highlights data gaps that must be addressed for equitable cancer prevention.