Fishes of various trophic levels play an important role in the stability and balance of aquatic ecosystems. Metal contaminants can impair the survival and population fitness of fish at elevated concentrations. When universal water quality criteria (WQC) of metals are adopted to protect different species in different geographic regions, they may not adequately protect all fish due to a lack of consideration for site-specific environmental conditions and species assemblages. Additionally, obtaining credible toxicity data for rare and endangered species is challenging. Therefore, this study aims to develop a robust, machine learning-based method to predict the toxicity of metals to various fish species, including rare and endangered species, and combine it with the non-parametric kernel density estimation of the species sensitivity distribution (NPKDE-SSD) model to derive site-specific WQC for better ecosystem protection. We show that this machine learning-based approach, with consideration of physicochemical properties of metals, hydrochemical conditions, biological characteristics of fishes, and metal toxicities, as well as their relationships, can well predict the toxicity of 19 metals to various fish species. The method is applied to derive site-specific WQC (based on the hazardous concentration of 5%) of these metals for the Eastern Plain lake region in China. The study provides a novel, alternative approach to supplement the insufficient toxicity information for site-specific WQC derivation and potentially improve the protection of fish species.
Keywords: NPKDE-SSD; machine learning; metals; prediction; rare and endangered fish; toxicity; water quality criteria.