• English 
  • Spanish 

Predicting toxicity from sediment chemistry using artificial neural networks: screening tools for sustainable sediment management

Sediments can act as sinks of multiple chemicals that accumulate over time, making risk assessment and sustainable management difficult and complex. The assessment of sediment quality should therefore be carried out through tiered decision-making frameworks in sequential steps of increasing complexity and cost. Therefore, the development of tools to link chemical concentrations of contaminants to the potential for observing toxicity is of great interest, since these tools can be very useful in initial assessment steps for an efficient allocation of limited resources; e.g., they can serve to identify uncontaminated sediments with low probability of toxicity, in which further testing would not be necessary, or to identify sediments that are highly likely to be toxic and would require attention in subsequent management steps.

Some existing approaches, such as the Logistic Regression Models (LRM)1 or the widely used empirical Sediment Quality Guidelines2-3, focus on estimating toxicity based on individual-contaminant models, which can be problematic since many chemicals occur together and the toxicity of the sediment is the product of complex interactions among them. The objective of this study is to present a new alternative approach in which advanced machine learning techniques based on Artificial Neural Networks (ANNs4) are used to model the probability of toxicity responses from sediment chemistry, considering the concentrations of the most commonly analyzed contaminants simultaneously and their interactions.

Available historical data from the U.S. EPA’s Environmental Monitoring and Assessment Program (EMAP) were compiled, and after thorough processes of data verification and screening, a large database of 1552 field-collected sediment samples with matching chemistry and toxicity data was prepared. The resulting database contained information on 9 metals (As, Cd, Cr, Cu, Pb, Hg, Ni, Ag, Zn) and 10 organic chemicals (low molecular weight PAHs, high molecular weight PAHs, PCBs, p,p’-DDD, p,p’-DDE, p,p’-DDT, TBT, hexachlorobenzene, alpha-chlordane and lindane), while sediment toxicity was evaluated in 10-day survival tests using the amphipod Ampelisca abdita.

Different data preprocessing techniques and different configurations of network architectures (Multi Layer Perceptrons) were tested performing cross-validation experiments, and the quality of the predictions was assessed using measures of resolution and reliability. The resulting models were compared to LRM obtaining a better classification performance both in terms of sensitivity (ability to correctly classify a toxic sample as toxic) and specificity (ability to correctly classify a nontoxic sample as nontoxic). Moreover, the economic value curves derived from the ANN models proved to be useful screening tools for sustainable sediment management, allowing decision makers to achieve the desired compromises between environmental protection (i.e. high sensitive) and efficient use of resources (i.e. high specificity to avoid wasting money and effort incorrectly considering nontoxic samples for further testing).

1 Field et al. 2002. Environ. Toxicol. Chem. 21: 1993-2002.
2 Wenning et al. 2005. Use of sediment quality guidelines and related tools for the assessment of contaminated sediments. SETAC.
3 Alvarez-Guerra et al. 2007. Integr. Environ. Assess. Manag. 3: 529-538.
4 Bishop, C.M. 1995. Neural Networks for Pattern Recognition. Oxford University Press.