Assessing the Interpretability–Performance Trade-Off of Artificial Neural Networks Using Sentinel Fish Health Data

Abstract

A number of sentinel species are regularly sampled from the environment near the Oil Sands Region (OSR) in Alberta, Canada. In particular, trout-perch are sampled as a proxy for the health of the aquatic ecosystem. As the development of the OSR began before the environmental monitoring program was in place, there is currently no established measure for the baseline health of the local ecosystem. A common solution is to calculate normal ranges for fish endpoints. Observations found to be outside the normal range are then flagged, alerting researchers to the potential presence of stressors in the local environment. The quality of the normal ranges is dependent on the accuracy of the estimates used to calculate them. This paper explores the use of neural networks and regularized regression for improving the prediction accuracy of fish endpoints. We also consider the trade-off between the prediction accuracy and interpretability of each model. We find that neural networks can provide increased prediction accuracy, but this improvement in accuracy may not be worth the loss in interpretability in some ecological studies. The elastic net offers both good prediction accuracy and interpretability, making it a safe choice for many ecological applications. A hybridized method combining both the neural network and elastic net offers high prediction accuracy as well as some interpretability, and therefore it is the recommended method for this application.