Spatial neighborhood effects and geographic embeddings in random forest modeling of mammography screening rates.

Abstract

e13643 Background: Mammography screening rates vary geographically across the United States, yet standard machine learning models often fail to account for spatial dependence, limiting their ability to resolve local screening disparities. This study examined whether incorporating spatial neighborhood effects and geographic embeddings improves random forest (RF) prediction accuracy and reduces residual spatial autocorrelation in tract-level screening estimates. Methods: PLACES 2024 tract-level mammography screening data were linked with socioeconomic and healthcare access indicators, including poverty, educational attainment, race, insurance coverage, median home value, proximity to mammography facilities, and the Social Vulnerability Index. Three RF regression models were trained using 5-fold cross-validation: (1) a base model with socioeconomic and access predictors; (2) a spatial-lag model including lag_mammo, defined as the mean screening rate of adjacent tracts; and (3) a spatial-embedded model adding geographic coordinate terms (longitude, latitude, longitude², latitude², and longitude×latitude). Model performance was evaluated using RMSE, MAE, R², and Moran’s I of residuals. Results: Incorporating spatial information improved performance across all metrics. The spatial-lag model reduced RMSE from 3.20 to 2.13 and increased R² from 0.55 to 0.80, eliminating residual spatial autocorrelation (Moran’s I = −0.01, p = 1.0). Adding coordinate embeddings further increased R² to 0.83. Conclusions: Incorporating spatial neighborhood effects and geographic embeddings substantially improved RF model performance and reduced spatial autocorrelation, supporting their use in identifying under-screened communities. This approach may assist in targeting interventions and informing strategies to reduce geographic disparities in preventive breast cancer care. Comparative predictive accuracy and spatial autocorrelation of three random forest models. Model RMSE MAE R² Moran’s I Base RF 3.20 2.50 0.55 +0.66 (p<0.01) Spatial-Lag RF 2.13 1.65 0.80 −0.01 (p=1.0) Spatial-Embedded RF 1.98 1.54 0.83 −0.046 (p=1.0) Negative Moran’s I values indicate no residual spatial dependence.