In brief
- An ecological analysis of 542 Medical Service Study Areas in California related age-adjusted lung cancer incidence for 2015–2019 to 27 community-level indicators of pollution, population characteristics and adult smoking prevalence.
- A support vector machine trained on 15 of those indicators separated high-incidence from low-incidence areas with a cross-validation AUC of 0.891, against 0.685 for a model using smoking prevalence alone.
- Smoking, linguistic isolation, neighbourhood socioeconomic status, unemployment and housing burden made the largest contributions to the classification.
- In regression, a higher neighbourhood socioeconomic status quintile (standardised β = −0.57) and a higher Hispanic population proportion (β = −0.51) were the strongest predictors of lower incidence.
- Discrimination was better for White-specific incidence (AUC 0.926) than for Hispanic-specific incidence (AUC 0.744), and the authors state that no causal interpretation is intended.
Study at a glance
- Design: Ecological analysis of area-level data
- Units: 542 Medical Service Study Areas in California
- Outcome: Age-adjusted lung cancer incidence rate, 2015–2019
- Indicators: 27 measures of pollution, population characteristics and adult smoking prevalence
- Best classifier: Support vector machine (radial basis function kernel) on 15 autoencoder-selected indicators, cross-validation AUC 0.891
- Comparator: Smoking-only model, AUC 0.685
- Strongest regression predictors of lower incidence: Neighbourhood socioeconomic status quintile (β = −0.57), Hispanic population proportion (β = −0.51)
- Subpopulation models: White-specific incidence AUC 0.926, Hispanic-specific incidence AUC 0.744
- Not reported: The size of the association with Black population proportion, which the abstract does not give
Fifteen community indicators discriminated better than smoking prevalence alone
The investigators matched incidence rates for 2015–2019 in each of the 542 study areas to 27 indicators, used an autoencoder to select 15 of them, and trained classifiers to distinguish areas of high incidence from areas of low incidence.
A support vector machine with a radial basis function kernel performed best, with a cross-validation AUC of 0.891. The model given smoking prevalence alone reached 0.685.
SHapley Additive exPlanations (SHAP) analysis named five indicators as the largest contributors to the classification: smoking, linguistic isolation, neighbourhood socioeconomic status, unemployment and housing burden.
Neighbourhood socioeconomic status gave the largest regression coefficient
Ordinary least squares regression returned a standardised β of −0.57 for neighbourhood socioeconomic status quintile and −0.51 for the proportion of Hispanic residents, each corresponding to lower age-adjusted incidence.
A higher proportion of Black residents was associated with higher incidence.
Weaker discrimination for Hispanic-specific incidence
Stratified by subpopulation, the models reached an AUC of 0.926 for White-specific incidence and 0.744 for Hispanic-specific incidence. The authors write that lower predictive performance in some subpopulations points to data gaps, which they say must be addressed for equitable cancer prevention.
“Community-level smoking, pollution, and population characteristics data improved the identification of regions with elevated lung cancer risk.”Pucillo et al., Cancer Epidemiology, Biomarkers & Prevention (2026)
Concurrent associations in one state, with no causal claim
The authors describe the associations as concurrent rather than aetiological. Because exposure and incidence were both measured for the area as a whole, an area with high incidence and high linguistic isolation does not show that its linguistically isolated residents carry the excess risk.
All 542 areas lie in California, and the abstract describes cross-validation within that dataset only, with no testing in another state or period. The authors propose that community-level indicators could identify areas of elevated lung cancer burden, including many with suppressed registry rates.
Sources
- Cancer Epidemiology Biomarkers & Prevention. Machine Learning Identifies Community-Level Factors Associated with Lung Cancer Incidence (2026-10-15). doi.org
Featuring Lung Summit faculty
This study was co-authored by Lung Summit faculty Sandip Patel.
This article was produced independently by the Lung Summit editorial team, without industry funding or input.
