Objective Cost-Informed Cutoff Criteria Improve the Utility of Machine Learning Models of Environmental Hazards: A Case Study of Groundwater Arsenic Distribution in India

Abstract

Although there are an increasing number of artificial intelligence/machine learning models of various hazardous chemicals (e.g. As, F, U, NO3−, radon) in environmental media (e.g. groundwater, soil), these most commonly use arbitrarily selected cutoff criteria to balance model specificity and sensitivity. This results in models of hazard distribution that, whilst often of considerable interest and utility, are not designed to optimize cost benefits of the mitigation of those hazards. In this case study, building upon recent machine learning modelling of the geographical distribution of groundwater arsenic in India, we show that the use of objective cost-informed criteria not only results in (i) different cutoff values for the classification of areas as of high or low groundwater arsenic hazard but also, more importantly, (ii) a reduction of overall potential (mitigation + testing + health impacts) costs. Further, we show that the change in optimal cutoff values and the reduction in overall costs vary from state to state depending upon locally specific classification-dependent costs, the prevalence of high arsenic groundwaters, the heterogeneity of the distribution of those high arsenic groundwaters, and the extent to which inhabitants are exposed to the hazard. It follows more generally that using cost-optimized criteria will result in different, more objective, and more cost-relevant appropriate balances being made between specificity and sensitivity in modelling environmental hazard distribution in different regions. This indicates also the utility of developing machine learning models at an appropriate local (e.g. country, state, district) scale rather than more global scales in order to better inform local-scale mitigation strategies.