Predicting Enteric Virus Presence in Surface Waters Using Artificial Neural Network Models

Many researchers have successfully used multivariate logistic regression models and artificial neural networks (ANN) for modeling environmental datasets when there are sufficient data observations to support their use. Recent advances in training methods of ANN models have allowed for the use of a greater number of observations of a limited dataset for model training, enhancing the usefulness of available data. This study employs an advanced relative strength effect (RSE)-based ANN training termination procedure for developing models for the prediction of the presence or absence of viruses in raw surface waters. The presence/absence of enteric viruses was predicted using a blend of input parameters that represented fecal load, age, and source as well as parameters that indicated the degree of change in watershed conditions. A conventionally trained ANN model was compared against an RSE-based trained ANN model and a multivariate logistic regression for prediction accuracy. The RSE-based ANN modeling method performed better predicting validation values than multivariate logistic regression. It predicted viral presence and absence with accuracies greater than 91 and 88%, respectively. The conventionally trained ANN model performed as well as the RSE-based ANN model in predicting overall virus presence (87.9%), and performed better in predicting overall virus absence (91 vs. 81%, respectively). Virus presence was strongly associated with the presence of epicoprostanol and AC/TC ratio values <15.