Spatiotemporal Machine Learning Integration of Atmospheric Reanalysis and Mammographic Data for Breast Lesion Malignancy Prediction

Air pollution has been investigated as a potential risk factor for breast cancer [1]; however, its quantitative impact on malignancy risk stratification remains uncertain, particularly when integrated with radiological features. In this study, we investigate whether long-term exposure to air pollution — a climate-sensitive environmental stressor — derived from Copernicus Atmosphere Monitoring Service (CAMS) reanalysis data provides complementary information for predicting breast lesion malignancy in a screened population.

We analysed mammographic and clinical data from 906 women undergoing breast cancer screening, classified as benign (BI-RADS B2) or malignant (BI-RADS B5). Individual exposure to NO₂, PM₂.₅, PM₁₀ and O₃ was estimated by linking the zip code of residence to CAMS gridded concentrations, computing both annual mean levels and cumulative exposure over the three years preceding diagnosis. Environmental exposure metrics were integrated with radiological descriptors, including lesion morphology, margins and breast density patterns, together with demographic information.

To reduce model complexity and limit overfitting, univariate feature selection was applied using an ANOVA F-test (p < 0.05) prior to training a feed-forward neural network. Model performance was assessed using independent validation data and compared with models excluding environmental exposure variables.

The integrated model achieved a ROC-AUC of 0.78, with balanced accuracy and a weighted F1-score of 0.73. Radiological features such as spiculated margins and irregular lesion shape remained the strongest predictors of malignancy; however, cumulative NO₂ and PM₂.₅ exposure metrics retained independent statistical significance and contributed to model performance. Limiting partially redundant air-quality metrics decreased apparent predictive power but improved model stability and interpretability, highlighting the potential impact of spatial and exposure-related confounding in observational datasets.

These findings suggest that long-term air-pollution exposure, as quantified using Copernicus atmospheric reanalysis products, provides a modest but consistent contribution to breast lesion malignancy risk stratification when combined with mammographic features [2]. This study demonstrates the feasibility of integrating atmospheric reanalysis data with clinical imaging information for exploratory environmental health applications, while underscoring the need for geographically robust validation and cautious interpretation of causality.

 

[1] White AJ, Bradshaw PT, Hamra GB. Air pollution and Breast Cancer: A Review. Curr Epidemiol Rep. 2018 Jun;5(2):92-100. doi: 10.1007/s40471-018-0143-2. Epub 2018 Mar 27.  

[2] Fiore, M.; Palella, M.; Ferroni, E.; Miligi, L.; Portaluri, M.; Marchese, C.A.; Mensi, C.; Civitelli, S.; Tanturri, G.; Mangia, C. Air Pollution and Breast Cancer Risk: An Umbrella Review. Environments 2025, 12, 289. https://doi.org/10.3390/environments12050153