Hybrid modeling and prediction of oyster norovirus outbreaks

Norovirus is a highly contagious virus with a low infectious dose, high infectivity, and efficient transmission in natural and human-made environments. Norovirus may concentrate in oysters, which are filter feeders and pump a large volume of water along with particles (such as norovirus) in water through their tissues if the oyster growing water is contaminated with sewage due to overland runoff, combined sewer overflow, and failing septic systems (Wang & Deng 2016). The consumption of raw oysters, therefore, may cause human norovirus infections and even large-scale outbreaks (Siebenga et al. 2009). Norovirus outbreaks due to the consumption of contaminated oysters resulted in closures of oyster harvest areas worldwide (David et al. 2007; McIntyre et al. 2017). For example, norovirus outbreaks in late 2016 and early 2017, in British Colombia, Canada, sickened more than 100 people and caused closures of 13 oyster farms (McIntyre et al. 2017). Consequently, it is vital to develop norovirus prediction models for informing decisions and implementing management interventions.

There is substantial evidence in the scientific literature that various environmental factors, such as water temperature, gage height, sunlight, salinity, wind, and precipitation, influence the survival and persistence of norovirus in the aquatic environment (Chenar & Deng 2017a, 2017b). While predicting viral contamination in the open natural environment (particularly coastal waters) is challenging due to unknown sources, complex marine environment (dilution and currents), and data scarcity (Pommepuy et al. 2005), machine learning-based models have shown promise in predicting norovirus outbreaks in oyster harvest areas (Wang & Deng 2016). Chenar & Deng (2018a) presented a Genetic Programming (GP)-based model for identifying the environmental conditions that trigger oyster norovirus outbreaks and predicting outbreaks in the Gulf of Mexico using the six environmental indicators, including water temperature, gage height, solar radiation, salinity, rainfall, and wind. Findings indicated that the GP-based model provided an efficient and effective tool for predicting potential oyster norovirus outbreaks. Chenar & Deng (2018b) also created an artificial intelligence (AI)-based forecasting model, called ANN-2Day model with a 2-day lead time, using the same six environmental predictors. The ANN-2Day model can forecast where, when, and under what environmental conditions oyster norovirus outbreaks occur. While the GP and ANN models are effective tools for predicting oyster norovirus outbreaks, applications of the models require daily data of the environmental parameters for all oysters harvest areas, making the in situ data collection challenging and even unlikely as the required data are not available daily for all oyster harvest areas along the U.S. Gulf coast. In fact, complete environmental datasets are available only in the areas highlighted in Supplementary Material, Figure S1, including Louisiana oyster harvest areas 1, 2, 3, 6, 7, 12, 13, 14, 15, 16, 18, 20, 24, 26, 27, 28, 29, 30, and Copano Bay and San Antonio Bay in Texas. The scarcity of in situ environmental data makes the in situ data-based management of oyster safety challenging.

Satellite remote sensing can be a useful technology for monitoring large areas like oyster harvest areas where field measurements are mostly not available. The data derived from NASA (National Aeronautics and Space Administration) EOS Terra/Aqua (MODIS (Moderate Resolution Imaging Spectroradiometer)) sensors (Tatem et al. 2004) can provide a robust and cost-effective approach for long-term monitoring of oyster harvest areas. Although norovirus cannot be sensed directly, remote-sensing data can be utilized to determine the environmental indicators of oyster norovirus. In fact, various MODIS spectral bands and their ratios have been widely used to quantify environmental indicators such as chlorophyll-a, colored dissolved organic matters (CDOM), Secchi disk depth, total phosphorus, water temperature, salinity, and gage height (Wang et al. 2005; Menken et al. 2006; Wu et al. 2009; Qing et al. 2013; Morozov et al. 2015). Wang & Deng (2017, 2018a, 2018b) presented a number of remote-sensing algorithms for retrieving sea surface salinity, sea surface temperature, and gage height using MODIS Aqua data, making it possible to predict oyster norovirus outbreaks using satellite remote-sensing data. In combination with an AI approach MODIS Aqua data, which reflect the characteristics of environmental parameters controlling norovirus epidemics in water, can be a useful data source for predicting human health risks associated with viral outbreaks.

The overall goal of this study is to demonstrate the efficacy of predicting oyster norovirus outbreaks by synergistically combining the AI-based modeling technique and satellite remote-sensing data directly from the MODIS sensor aboard the NASA Aqua spacecraft. To that end, a hybrid Principle Component Analysis (PCA) and Artificial Neural Network (ANN) modeling approach was developed in this study to establish a predictive relationship between oyster norovirus outbreaks and MODIS Aqua data from ocean color bands.

To assess the suitability of the data for PCA analysis, Kaiser–Meyer–Olkin (KMO) and Bartlett's tests were performed. The KMO index is a measure of sampling adequacy varying from 0 to 1 with values greater than 0.70 being considered suitable for PCA analysis (Budaev 2010). Bartlett's test of sphericity checks if there is a redundancy between variables that can be summarized with some components. When the test is significant (p < 0.05), it can be assumed that the correlation matrix is significantly different from an identity matrix and PCA can be applied (Williams et al. 2010). Cattell's screen test was used to determine the number of factors to retain. The test involves a visual examination of the eigenvalue plot in which only the components, forming the elbow before the slope of the graph goes from steep to flat, are kept (Budaev 2010). In addition, Varimax orthogonal rotation, which is the most popular rotation method (Kaiser 1958), was specified in this study. The statistical software package IBM SPSS was used to implement this technique.

The AI-based ANNs method has been increasingly applied in classification, clustering, prediction, and many other areas (Du 2010; Khashei & Bijari 2012; Castellani 2013; Wang & Deng 2016, 2018a, 2018b; Chenar & Deng 2017). ANN has also been proven to be an effective tool for describing nonlinear relationships between norovirus outbreaks and environmental variables in coastal waters (Wang & Deng 2016; Chenar & Deng 2017). In this study, a three-layer feed-forward neural network with back-propagation learning was constructed to develop a model for predicting oyster norovirus outbreaks using MODIS data. The architecture of the ANN model includes an input layer (receiving model input data), a single hidden layer (consisting of 20 neurons), and an output layer (showing the norovirus outbreak risk in any oyster harvest area for a given day). The number of hidden neurons depends on the training pattern, and there is no general rule for it in ANN (Ahmad et al. 2017). The neuron number of 20 was selected through a trial and error procedure by the re-training model using 5, 10, 15, 20, and more neurons. It was found that the neuron numbers higher than 20 would not significantly improve the model performance in terms of true positives, false positive, false negative, and true negative rates. Therefore, 20 neurons were chosen to reduce the model complexity. The ANN model training and testing processes were performed using the Neural Network Toolbox in the MATLAB program. The ANN model was created using three subsets of the available data, including the training dataset to adjust the weights, the validation dataset to measure network generalization, and the testing dataset to assess the performance. This data-partitioning step was conducted to evaluate the ability of the model to generalize through the comparison of predictions with the remaining data that were not used in the training process. Specifically, the normalized datasets from 2007 to 2010 were randomly split into three groups for training (accounting for 60% of the datasets), validation (20%), and testing (20%). Random data division improves model performance and prevents overfitting (Palani et al. 2008). The best-trained ANN model was identified based on the performance of top-ranked models in forecasting oyster norovirus outbreaks. The ANN model predictions were then compared with confirmed historical norovirus outbreaks to determine a threshold value for model-predicted norovirus outbreak risks that were consistently associated with confirmed outbreaks. Moreover, the cross-validation was performed to assess the predictive ability of the model using the independent data collected from 2011 to 2016 that were excluded from the model development phase. Finally, the overall performance of the model was evaluated using true positive and true negative rates.

A hybrid PCA-ANN prediction model was created by using the outputs of candidate ANN models as inputs of the optimum PCA model that accurately predicted all historical oyster norovirus outbreaks with a minimum number of false outbreaks. Specifically, the 15 PCs were initially used as input variables of ANN models, producing 100 candidate ANN models (outputs Y1–Y100) (Supplementary Material, Figure S2). The 100 initially trained ANN models were then used as inputs to generate 50 refined ANN models (Z1–Z50). Finally, the 50 refined ANN models were utilized as the model input variables to develop the hybrid PCA-ANN model for predicting norovirus outbreaks (NOV). Supplementary Material, Figure S2 summarizes the steps involved in the development of the hybrid PCA-ANN prediction model.

The overall performance of the hybrid PCA-ANN model was evaluated using a confusion matrix. The number of days with reported outbreaks and the number of days without reported outbreaks were counted from 2007 to 2016. The days, on which the model predicted risk higher than the threshold but no outbreaks were reported, were labeled as false positives and the days, on which norovirus outbreaks were reported but the model predicted risk lower than the threshold, were treated as false negatives. Finally, true positive and negative rates and positive and negative predictive values were calculated.

Supplementary Material, Table S3 summarizes the results from the KMO and Bartlett tests. It can be seen from the table that the KMO statistic is 0.88 (close to 1), confirming the capability of using the PCA method in reducing the number of input variables. According to Bartlett's test, the significance level is 0 (p < 0.05), indicating that the dataset is suitable for reduction and PCA can be useful for the dataset. The first 15 PCs were selected as inputs of the ANN model based on Cattell's screen test (Figure 1) and the cumulative variance proportion explained by the components. In fact, the interpretation of a screen test plot is subjective, requiring the judgment of the researcher to select the number of optimum PCs in such a way that it would adequately describe the input variable characteristics (Thompson 2004; Tabachnick & Fidell 2007). Table 2 presents eigenvalues, variance proportion, and cumulative variance proportion for the first 15 PCs. According to the table, it is clear that the first 15 PCs (PC1–PC15) represent 96.59% of the total variance proportion of input variables. According to the eigenvector values, the band ratios, involving spectral bands 8, 9, and 13, have the significant effects on the PC1 that described more than 53% of the variance proportion of the input variables. According to Table 1, bands 8, 9, and 13 were utilized to retrieve salinity and gage height from MODIS satellite data. Furthermore, band 9, which has the most strong impact on the second component (PC2), explains more than 15% of the variable variance. Likewise, PC3 is affected by the normalized fluorescence line-height, which could be considered as a solar radiation indicator, as Normalized Fluorescence Line-Height is a measure of the solar stimulated chlorophyll-a fluorescence.

The hybrid PCA-ANN model was selected from the top five ranked candidate models in terms of true positive and negative rates. A model-based threshold risk of 0.5, which consistently predicted the reported outbreaks, was selected by comparing predicted risks of norovirus outbreaks based on the ANN model with the occurrence of observed epidemics. Figure 2 shows the comparison between the PCA-ANN model-predicted risks of norovirus outbreaks and the observed norovirus outbreaks in oyster harvest areas along the U.S. Gulf of Mexico coast.

The PCA-ANN model predicted two clusters of norovirus outbreaks in Louisiana Area 3 for 12 December 2007 and 3 March 2010 with risks of 1 and 0.69, respectively. Furthermore, Figure 2 shows daily risks of norovirus outbreaks in the San Antonio Bay oyster harvest area where two oyster norovirus outbreaks were reported for 1–24 February 2007 and 16–25 November 2009, respectively. The PCA-ANN model was capable of reproducing the first confirmed outbreaks for 7, 16, and 18 February 2007 with the risk of 1, but the model did not predict the second reported outbreak in November 2009. The PCA-ANN model was capable of reproducing one out of the two confirmed outbreaks for January 8 with the risk of 0.71 in the Mississippi oyster harvesting Area IIC where two oyster norovirus outbreaks were reported in January and February 2009 (Supplementary Material, Table S2). Figure 2 demonstrates that the PCA-ANN model predicted multiple norovirus outbreaks for 6, 18, and 25 March 2010 with the risks higher than the threshold of 0.5, which were consistent with the inferred norovirus outbreak period of 6 and 24 March 2010 in Louisiana Area 7. There were no reported norovirus outbreaks from 2007 to 2010 in areas 1, 2, 10, 11, 12, 14, 15, 19, 21, 24, 25, 28, and 29 and the PCA-ANN model did not produce any false outbreaks for those areas (Figure 2).

To validate the performance of the PCA-ANN model, the additional 6 years of environmental and epidemiological data from 2011 to 2016, which were excluded from the model development, were employed for cross-validation of the model. Figure 3 illustrates the cross-validation results of the PCA-ANN model with the 6 years of independent data collected from Louisiana Areas 23 and 30 and Texas oyster harvest area in Copano Bay. The PCA-ANN model predicted the risk of 0.76 for 24 April 2012, during the reported outbreak period, as illustrated in Figure 3 for Louisiana Area 23. It can be seen from Figure 3 that the model did not predict the reported norovirus outbreak in Area 30 that was closed on 4 January 2013 after 12 people were infected by norovirus due to eating raw oysters harvested from this area between 28 December 2012 and 4 January 2013. There was a reported norovirus outbreak between 26 December 2013 and 9 January 2014 in Copano Bay, Texas. The PCA-ANN model predicted a norovirus outbreak with the risk of 0.61 for 3 January 2014 during the reported outbreak period, as shown in Figure 3. The PCA-ANN model also predicted two unconfirmed norovirus outbreaks for this area for 2 November 2015 and 13 January 2016. Therefore, these unconfirmed norovirus outbreaks could be a true outbreak that was not reported or a false cluster of outbreaks due to the absence of viruses or a source of fecal contamination in the environment. The model was further validated for areas 1, 2, 10, 11, 12, 14, 15, 19, 21, 24, 25, 28, and 29 where no confirmed outbreaks occurred from 2011 to 2016. The PCA-ANN model did not produce any false outbreaks for these areas.

In terms of the model performance, the total number of true positives, false positive, false negative, and true negative were 8, 1, 3, and 10,304, respectively. The true positive and negative rates of model predictions were 72.7 and 99.9%, and positive and negative predictive values were 88.9 and 99.9%, respectively, demonstrating the ability of the model to predict the outbreaks. The relatively low true positive rate may be caused by the fact that some important environmental predictors, such as rainfall and wind (Chenar & Deng 2018a, 2018b), are not included in the new model.

While an oyster norovirus outbreak may theoretically occur at any risk level, our findings, which are based on a comparison of the hybrid PCA-ANN model predictions with reported norovirus outbreaks, indicate that reported historical oyster norovirus outbreaks in the human population were consistently associated with the risk range of 0.5–1.0. Therefore, the risk of 0.5 was selected as the risk threshold value for oyster norovirus outbreaks. It should be noted that the oyster norovirus outbreaks predicted with the hybrid PCA-ANN model only refer to the norovirus outbreaks associated with the consumption of contaminated oysters. The predicted oyster norovirus outbreaks do not include the human norovirus outbreaks caused by the spread of norovirus from person to person. It should also be pointed out that a lower norovirus outbreak risk (such as 0.55) predicted with the hybrid PCA-ANN model might cause a lower number of human infections. However, a quantitative relationship between the model-predicted outbreak risk and the number of human infections is not available at this time due to the lack of detailed data for the norovirus concentration in oysters and associated epidemiological data. As a result, a model prediction can only be interpreted as an outbreak if the predicted risk is in the range of 0.50–1.00 or no outbreak if the predicted risk is in the range of 0.00–0.49. More efforts are needed to link a model-predicted risk level to a human infection level.

The application of the PCA method reduced the number of input variables of the ANN model from 65 to 15 without eliminating useful information, reducing model training time, and improving model performance. The majority of remote-sensing band ratios were loaded on the first component, demonstrating the efficiency of using band ratios in the model development. The PCA method was effective in terms of reducing and identifying the input variables for the ANN model due to the lack of knowledge of the nonlinear relationship between an oyster norovirus outbreak and its environmental drivers described with MODIS band data in this study. Further studies in close consultations with norovirus epidemiologists and caseworkers are needed to improve understanding of the correlation between remotely sensed data and oyster norovirus outbreaks in coastal waters.

As compared with the nowcasting and forecasting models developed previously by the authors' research group (Wang & Deng 2016; Chenar & Deng 2018a, 2018b), a unique advantage and important feature of the hybrid PCA-ANN remote-sensing-based model is its capability of making daily prediction of the oyster norovirus outbreak risks for all oyster harvesting areas along the U.S. Gulf coast as long as the MODIS ocean color data are available, enabling managers to make timely decisions on oyster safety due to the expansion of the spatial coverage and temporal frequency of oyster safety monitoring. Specifically, the hybrid PCA-ANN model can be employed to predict norovirus outbreaks as long as MODIS Aqua satellite data are available. The PCA-ANN model expands the spatial coverage of oyster safety monitoring to larger geographical areas where complete in situ environmental datasets are not available. Due to the daily prediction, the PCA-ANN model also serves as an early warning system since oyster norovirus outbreaks in the human population generally occur 1–2 weeks (including the incubation period of 12–48 h for norovirus-associated gastroenteritis) after norovirus-contaminated oysters are harvested (Wang & Deng 2016). Hence, the hybrid PCA-ANN and remote-sensing-based model can be used as an effective modeling framework for improving the management of oyster safety and public health by ensuring that oysters are safe to harvest. Figure 4 shows how the remote-sensing data-based PCA-ANN model could be applied in combination with other in situ data-based models, developed by the authors' research group, for prediction of oyster norovirus outbreaks, and how the model predictions could be utilized to implement management interventions, such as field sampling, laboratory analysis, and oyster bed closures, for protecting public health while minimizing oyster ground closures.

Since a model may produce false predictions or alerts, as shown in Figures 2 and 3 (particularly San Antonio Bay in December 2009, Mississippi Area IIC in March 2009, and Louisiana Area 30 in December 2012), and the false alert-based closure of an oyster harvest area might cause substantial economic damage to oyster farmers, a sound-science-based procedure (Figure 4) should be followed in using the model predictions. Depending on the type of available data, the PCA-ANN model or the models based on in situ measurement of environmental data could be run to predict daily risks of oyster norovirus outbreaks. If a model-predicted risk of norovirus outbreak exceeds a predefined outbreak risk threshold, oyster harvesting in the implicated area should be temporarily suspended and sampling should be conducted to confirm or deny the model prediction. Specifically, the following procedure should be followed, whenever a model-predicted risk of norovirus outbreak exceeds a predefined outbreak risk threshold to prevent oyster norovirus outbreaks:

1. Water and oyster samples should be collected from the oyster harvest areas where the model predicts a potential outbreak;

2. Laboratory analyses of the water and oyster samples should be conducted to verify the presence or absence of norovirus in oysters; and

3. Implicated harvest areas should be closed if the laboratory results confirm the presence of norovirus in oysters.

This is the first major study on predicting potential oyster norovirus contamination in coastal waters by using satellite remote-sensing data. While this study focused on the oyster harvest areas along the U.S. Gulf of Mexico coast, the remote-sensing-based hybrid modeling approach could be potentially applied to other regions and countries due to the global coverage of the NASA MODIS satellite data. Therefore, this study could be expanded in the future toward the development of a global-scale hybrid modeling system for providing daily predictions of norovirus outbreak risks in any oyster harvest areas, serving as a global-scale early alert system for potential oyster norovirus outbreaks and thus protecting public health. The model enables oyster management authorities to expand the monitoring and prediction of norovirus outbreak risks from areas where monitoring data are accessible to other oyster harvest areas where monitoring stations are not available. Although this study provided convincing evidence that predicting oyster norovirus outbreaks using satellite remote-sensing data are possible, field sampling-based validation of model prediction is always needed. A major limitation of the PCA-ANN model application is that the model may not work during the cloud cover days when the remote-sensing data, specifically MODIS Aqua data, are not available. Also, it appears from Figures 2 and 3 that the PCA-ANN model works well for Louisiana oyster harvest areas as the model was developed primarily using the data from the Louisiana oyster harvest areas. However, the model tends to produce more false predictions for other oyster harvest areas beyond Louisiana, although the model still works for the other areas with a lower efficacy. It is, therefore, recommended that the modeling and intervention procedure shown in Figure 4 be followed, by using both the in situ data and the remote-sensing data, to protect public health even during cloudy days. While the application of the model to other norovirus concerned oyster harvest areas like those in British Columbia may produce a higher rate of false predictions and thus require additional testing of the model with local data, the modeling methodology can be adopted for other norovirus concerned oyster farms to monitor any potential oyster safety threats.

This paper presents a hybrid monitoring and modeling approach for monitoring and predicting potential risks of oyster norovirus outbreaks in the human population. In terms of monitoring, high-resolution satellite remote-sensing data covering all oyster harvesting areas were utilized to provide daily data needed in modeling. In terms of modeling, a hybrid PCA and ANN model was created for the daily prediction of oyster norovirus outbreak risks and demonstrated using the data from the Northern Gulf of Mexico that is prone to epidemics. Based on the results of this paper, the following conclusions can be drawn:

• A remote-sensing-based PCA-ANN model is an effective tool for predicting oyster norovirus outbreaks. Specifically, the PCA-ANN model achieved the true positive and negative rates of 72.7 and 99.9%, respectively, in predicting reported oyster norovirus outbreaks, demonstrating the efficacy of the model in predicting outbreaks, and minimizing the adverse impact of outbreaks on human health and the shellfish industry.

• The PCA-ANN model greatly expands the spatial coverage of oyster safety monitoring programs from limited areas where in situ monitoring data are available to all oyster harvest areas including areas where in situ monitoring data are not available.

• The PCA-ANN model also greatly expands the temporal frequency of water quality monitoring regularly conducted by oyster safety monitoring programs from 1 month to 1 day, enabling oyster safety monitoring programs to make daily decisions on oyster safety and respond effectively to any potential oyster safety threats by implementing management interventions.

• Field validation is always needed, particularly when the PCA-ANN model predicts a high risk of an oyster norovirus outbreak, to confirm the presence of model-predicted norovirus outbreak in oyster harvest areas.

The material is based upon work supported by the U.S. NASA (National Aeronautics and Space Administration: award No. 80NSSC20M0216) and the Louisiana Board of Regents (LEQSF(2020-23)-Phase3-14). We appreciate the insightful comments from anonymous reviewers and particularly those from Charmaine Enns, Eleni Galanis, Natalie Prystajecky, Jackie Plamondon, Lorraine McIntyre, and Theresa Burns of the British Columbia Center for Disease Control and Prevention, Canada, illuminating the final form of this paper.

All relevant data are included in the paper or its Supplementary Information.