Machine learning prediction of community-level lung cancer incidence in California

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

Background 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. Methods 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…

By the Truvace deskOctober 20, 20262 min read1 sourcePeer-reviewedTRV-2026-1985· cite & verify

Machine Learning Identifies Community-Level Factors Associated with Lung Cancer Incidence — Adenocarcinoma - CT scan (5499628365) by Yale Rosen from USA. CC BY-SA 2.0 · https://creativecommons.org/licenses/by-sa/2.0

Trace impact readingPositive state

P 62

Both sides are scored from claims and sources, not community votes.

G 68

In brief

Researchers analyzed 542 Medical Service Study Areas in California, linking lung cancer incidence rates from 2015 to 2019 with 27 indicators covering pollution, population characteristics, and adult smoking prevalence. They trained machine learning models to separate high-incidence areas from low-incidence ones and used regression to measure how each factor related to incidence.

The findings suggest area-level data can help find places with elevated lung cancer burden, including places where registry rates are suppressed, which may inform prevention planning. The authors stress that the associations are concurrent rather than causal, and weaker performance for some subpopulations means the approach may serve some communities less well.

Main points

The best model was a Support Vector Machine with an radial basis function kernel trained on 15 autoencoder-selected indicators, reaching a cross-validation AUC of 0.891.
A smoking-only model reached an AUC of 0.685, so adding pollution and population data improved identification.
SHAP analysis ranked smoking, linguistic isolation, neighborhood socioeconomic status, unemployment, and housing burden as top contributors.
In regression, neighborhood SES quintile and Hispanic population proportion were the strongest predictors of lower incidence, while a higher Black population proportion was linked to higher incidence.

The gain

A machine learning model using community indicators identified high-incidence California areas better than smoking data alone, which could help locate areas of elevated lung cancer burden where registry rates are suppressed.

The problem

The model predicted Hispanic-specific incidence less accurately than White-specific incidence, exposing data gaps that could undermine equitable cancer prevention.

The rundown

The study was an ecological analysis at the level of Medical Service Study Areas rather than individuals. It used age-adjusted lung cancer incidence rates from 2015 to 2019 and 27 community indicators, with SHAP used to interpret which indicators drove the classification.

Ordinary least squares regression quantified associations: neighborhood SES quintile had a standardized beta of -0.57 and Hispanic population proportion had a beta of -0.51 as predictors of lower age-adjusted incidence.

Stratified models by subpopulation showed a performance gap, with an AUC of 0.926 for White-specific incidence versus 0.744 for Hispanic-specific incidence. The authors say this points to data gaps that need to be addressed for equitable prevention.

What this doesn’t fix

The authors state the associations are concurrent rather than etiologic and not causal, and predictive performance was lower for some subpopulations, indicating data gaps.