Modeling and forecasting oyster norovirus outbreaks of long-range dependence on antecedent marine environmental conditions.

This study presents three artificial intelligence-based models - XGBoost, Random Forest (RF), and Deep Artificial Neural Network (DANN) - with 2-day lead time for forecasting broad-scale oyster norovirus outbreaks. Among them, the XGBoost model performs best and is characterized by unique features: (1) the model was constructed and tested with epidemiological and environmental data collected from three distinct coastal countries representing broad trends in oyster norovirus outbreaks; (2) model input variables include five important environmental predictors (along with their time-lags ranging from 2-30 days), identified by using Pearson correlation and Gini index, including solar radiation, water temperature, gage height, precipitation, and salinity plus GPS coordinates, which affect various processes governing the sources and sinks of oyster norovirus; and (3) the model output performance is characterized by a high accuracy, demonstrating the strongest overall prediction performance. This study identifies solar radiation and water temperature as the two most important predictors of oyster norovirus outbreaks. Another major finding is the strong long-range dependence of outbreaks on antecedent marine environmental conditions up to 30 days before. These novel features of the XGBoost model provide an effective early-warning framework that can support proactive management of oyster norovirus outbreaks.