Spatially explicit non-linear models for explaining the occurrence of infectious zoonotic diseases

Zoonotic diseases can be transmitted via an arthropod vector and disease risk maps are often based on underlying associative factors within the surrounding landscape of known occurrences. A major limitation is the ability to track disease incidence at a meaningful geographic scale and traditional linear modeling approaches may not always be appropriate. The objective is to create meaningful geospatial risk models at the ZIP code level describing the occurrence of 2 tick-borne zoonotic diseases (Lyme disease [LD] and Rocky Mountain spotted fever [RMSF]) known to occur in Tennessee. Case claims with ICD-9 diagnosis codes for LD (088.81) and RMSF (082.0) were extracted from a managed care organization data warehouse. Four separate modeling techniques were constructed (logistic regression, classification and regression tree [CART], gradient boosted tree [GBT], neural network [NNET]) and compared for accuracy. Post-analysis predictive risk maps were created using a geographic information system (GIS) and compared to simple incidence risk maps. Areas higher in disease prevalence were not necessarily the same areas having high predicted disease risk. Non-linear modeling provided better results than traditional regression-based approaches. GBT explained LD incidence (misclassification rate: 0.232; ROC: 0.789) and NNET explained RMSF incidence (misclassification rate: 0.288; ROC: 0.696). Covariates explaining disease risk included forested/non-forested wetland area, urbanization and median income levels. Administrative medical claims data is not being leveraged in surveillance efforts, but is an inexpensive volume rich dataset for mapping disease risk and studying zoonotic infections. We recommend integrating administrative data with state registries to improve geospatial analytical capabilities.