The rapid expansion of machine learning applications in heavy metal risk research has generated a large but fragmented body of literature, necessitating a systematic summary of various methodologies. This survey examines 182 research articles from Web of Science, Scopus, and IEEE Xplore over the past 10 years, providing a comprehensive analysis of the application of prevalent machine learning algorithms across research areas related to heavy metal risks, covering more than 30 algorithms and 9 research areas. The results show that regional risk assessment, risk source analysis, risk driver analysis, and research on the pathogenicity of HMs are the four most frequently applied areas of machine learning, accounting for 80.06% of all applications. Classical machine learning, especially a series of tree-based algorithms, dominates across all applications, accounting for 69.93% and 42.74%, respectively. In addition, some auxiliary algorithms, particularly those for feature analysis, are often used in conjunction with machine learning, primarily to interpret model predictions and analyze risk sources or drivers. Three principal methodological challenges emerge from our review: (1) poor model generalizability across different environmental conditions; (2) insufficient reliability of predictions; and (3) difficulty in obtaining high-quality training data. Some current literature is also constrained by small sample sizes, regional bias, limited field validation, and insufficient integration of multi-omics data with machine learning pipelines. To address these gaps, we advocate for the routine adoption of explainable AI techniques with rigorous stability checks, the development of publicly available, field-validated benchmark datasets, and greater integration of mechanistic knowledge with data-driven models.

Effect-based toxicity assessments are crucial for evaluating the risks of disinfection byproducts (DBPs), particularly unknown species, generated during drinking water chlorination. However, the accuracy of this approach is highly dependent on unbiased sample extraction. Conventional methods often employ single, low-pH extraction, which may fail to recover pH-sensitive amphoteric DBPs derived from amphoteric precursors (e.g., nitrogenous compounds). This study investigated how extraction pH affects the measured biotoxicity of DBPs formed from three model precursors: biopterin (Bip), cytosine (Cyt), and tryptophan (Trp). Under excess chlorine conditions, all three precursors degraded rapidly. The formation of aliphatic DBPs followed the order Trp > Cyt > Bip, and the maximum toxicity of the non-volatile extracts, assessed via a Vibrio fischeri bioassay, followed the reverse order: Bip > Trp > Cyt. This toxicity profile was significantly influenced by extraction pH, with maximum toxicity observed for Bip at around pH 4.0, under weakly acidic conditions for Trp, and under neutral to alkaline conditions for Cyt. For all precursors, the total organic carbon concentration remained constant throughout chlorination, indicating negligible mineralization and the predominant formation of non-aliphatic, likely heteroaromatic, products. These findings demonstrate that conventional extractions at a single low pH can lead to the incomplete recovery of toxic DBPs from amphoteric precursors. Therefore, pH-optimized extraction protocols are necessary for a more accurate risk assessment of chlorinated drinking water.

Anomaly detection in water-related time-series data often reveals important environmental problems and serves as a starting point for scientific discoveries. Machine learning has become the mainstream method and a research hotspot for anomaly detection in recent years. This review examines 106 research articles from the Web of Science database published over the past 10 years. Unlike other surveys, this review focuses on anomalies arising from the water-related variables themselves rather than from equipment malfunctions. The work assesses the overall trends in the application and development of machine learning models for water-related anomaly detection. It classifies machine learning-based anomaly-detection models from two dimensions: development stage and anomaly-detection paradigm. Our analysis covers the mechanisms, strengths, limitations, and applications of various machine learning-based anomaly-detection models across different paradigms, highlighting current challenges and prospective research directions in water-related anomaly detection.