The aim of this study was to develop a biological early warning system for the detection of aquatic toxicity and test it with five toxicants with distinct chemical nature. This was done in order to verify the spectrum of sensitivities of the proposed system, as well as the potential identification capability of the tested contaminants, using only the analysis of zebrafish's behavior. Six experimental conditions were tested: negative control and five toxicants (bleach, lindane, tributyltin, mercury, and formaldehyde). The exposure time was 45 min, and the concentrations used corresponded to 9% of LC50's-96 h for the tested compounds, to ensure ecologically relevant results. A total of 108 fish were used, with each individual experimental condition being tested 18 times. A statistical model of diagnosis was used, combining self-organizing map and correspondence analysis. The values of sensitivity, specificity, accuracy, false positive, false negative, positive predictive value (PPV), and negative predictive value (NPV) were calculated. The objectives of the work were accomplished and the system showed a good overall diagnostic performance with 79% in accuracy, 77% in sensitivity, and 88% in specificity. The lowest result of the predictive values was 78% (lindane and mercury), in the case of the NPV, and 86% (bleach and lindane), in the case of the PPV. The best result of the predictive values was 100% (bleach and tributyltin), for the NPV, and 89% (tributyltin), for the PPV. Regarding the five tested toxicants, the system was able to correctly identify the agent responsible for the contamination in 40% of the positive diagnoses.
Keywords: Biological early warning system (BEWS); Neural network; Self-organizing map; Toxicant identification; Video tracking; Zebrafish (Danio rerio).