Application of neural networks to quantify the utility of indices of biotic integrity for biological monitoring

Indices of Biotic Integrity (IBIs) or multimetric indices have been developed as an approach for monitoring and evaluating biological condition of aquatic organisms. Quantitative evaluations of IBIs to determine whether they can explicitly link environmental condition with anthropogenic activities are needed to effectively use them in management. Analytical approaches using supervised neural networks are potentially powerful techniques to evaluate IBIs. The goal of this study was to evaluate the use of neural networks to identify ecosystem characteristics related to IBI response and to explicitly quantify relationships between variables using sensitivity analyses. An aquatic macrophyte-based IBI developed for Minnesota lakes was used as an example. The study was particularly interested in the usefulness of neural networks to highlight key predictors of IBI performance and to be used as a technique to evaluate multimetric index performance in other systems or regions. Neural networks made accurate predictions of overall IBI scores using an independent dataset, whereas predictive performance of the models varied for individual metrics. Bootstrap analyses to evaluate the effects of different training data on model performance indicated that predictions were highly sensitive to the training data. More conventional modeling techniques, such as multiple regression, performed similarly in predicting IBI scores, although diagnostic tools developed for neural networks provided novel insight into variables influencing IBI response. We suggest that neural networks have the ability to quantify ecological relationships that affect biotic integrity, but the statistical uncertainty associated with multimetric indices may limit the use of predictive models to infer causation. Accordingly, the statistical properties of multimetric indices should be carefully evaluated during index development, with specific attention given to the diagnostic capabilities of individual metrics.