Background: One-quarter of SARS-COV-2 (aka COVID-19) confirmed cases and deaths globally are in the United States (US), underscoring the need for a rigorous predictive model to inform healthcare decision making. Various forecast models have been developed, but very few use machine learning which typically offers greater predictive accuracy than traditional approaches.

Objective: To develop a proof-of-concept dynamic geospatial model for predicting positive new SARS-COV-2 cases in US states using the transparent Bayesian networks machine learning approach.

Methods: Targeted literature reviews were used to identify important predictive variables for positive new SARS-COV-2 cases. State-level data specifying identified variables were pooled from public sources, and final variable selection was informed by principal component analyses. A Bayesian network machine learning approach was used to identify interdependencies between variables. The outcome of interest was the predicted number of new positive SARS-COV-2 cases in each US state and county in 28 days from an index date. The model was trained with data from 40 randomly selected states and validated by K-fold cross validation. Model goodness-of-fit was assessed using visual inspection, AIC and Bayesian information criterion (BIC). Measures of predictive accuracy included root mean square error (RMSE), mean percentage error (MPE), mean absolute percentage error (MAPE) and area under the receiver operating characteristic curve (AUC - using a discretized outcome).

Results: Predictions from the dynamic Bayesian model were a close fit to observed data for all states. The ME, RMSE, MAE, MPE and MAPE were 61.5, 497.9, 320.4, 36.0 and 49.1 respectively.

Conclusion: We developed a geospatial dynamic Bayesian network model that accurately predicts positive new SARS-COV-2 cases in 28 days into the future for US states. The model’s prediction accuracy appears to be at least on par with other popular models that are available for public use.