A Bayesian Spatial Deep Learning Framework for Mapping ZIP Code-Level Syphilis Hotspots in Hillsborough County, Florida

Syphilis has re-emerged as a major public health concern in the United States, exhibiting substantial spatial heterogeneity that is inadequately captured by conventional surveillance systems. This study presents an integrated spatial-statistical and deep learning framework to identify ZIP Code Tabulation Area (ZCTA)-level syphilis hotspots in Hillsborough County, Florida, by combining epidemiological surveillance data, socioeconomic and racial covariates, and Sentinel-2 satellite-derived environmental indicators. Syphilis incidence was modeled using a Poisson generalized linear model with population offsets, while spatial dependence was addressed using a Bayesian hierarchical Conditional Autoregressive (CAR) model within the Besag-York-Mollie (BYM) framework. We further introduce an eigen-Bayesian convolutional neural network (EBCNN) that incorporated eigen-decomposition of spatial weight covariance matrices, enabling spectral regularization and improved representation of latent spatial autocorrelation in the upscaled sampled ZCTA georeferenced capture point data. Sentinel-2 Level-2A imagery was used to derive environmental predictors, including the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Built-up Index (NDBI), aggregated to ZCTAs through zonal statistics. Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting Tree classifiers were trained to distinguish high- and low-risk ZCTAs. Model performance was evaluated using cross-validation, area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. Results indicate that higher syphilis incidence is significantly associated with socioeconomic deprivation, increased urban density, and reduced vegetation cover. Incorporating spatial random effects reduced residual spatial autocorrelation and improved hotspot stability in ZIP Codes. Global Moran's I = 0.34, p < 0.001. Following implementation of the Besag-York-Mollie (BYM) spatial model, Moran's I decreased to 0.07 ( p = 0.18). The proposed EBCNN achieved superior hotspot detection accuracy and uncertainty calibration compared with conventional deep learning models. Integrating satellite-derived environmental variables with demographic and socioeconomic factors provides a scalable, high-resolution framework for precision public health surveillance and targeting potential syphilis ZCTA vulnerable populations for interventions.