Community Research and Development Information Service - CORDIS

FP5

CODTRACE Report Summary

Project ID: Q5RS-2001-01697
Funded under: FP5-LIFE QUALITY
Country: Spain

Deliverable: parasite community classification of harvest location of cod

As expected, cod from farms could not be allocated to harvest location based on studies of parasite assemblages. Due to artificial feeding these cod are seldom infected with parasites transmitted through the food web. However, parasite evidence can still be used to discriminate cod from farms or wild locations, since few cod from farms were infected with parasites, whereas only two wild cod (from the Baltic) were devoid of parasites.

Although 61 parasite species were found, a number between 10 and 15 species seemed reasonable in terms of cost-benefit to allocate cod to five or six harvest locations since the addition of more species to the models did not improve allocation results substantially. The suite of predictor parasite species varied between datasets, but eight species were selected in all of them, and one additional species was chosen in three of the four datasets. These nine species are helminths (two trematodes, six nematodes and one acanthocephalan), which are commonly found in either the digestive tract or the visceral cavity of cod in European waters. None of the species with prevalence equal or less than 10% was useful as predictor of harvest location. In practical terms, this means that focussing on collecting relatively common species is enough to develop methods to establish harvest location of cod in European waters.

The analyses suggested better classification ability of both predictive models (BBP network and LDA) when the season samples were analysed separately, since their performance increased correct classification rates of cod to harvest location by about 5%. This result suggests that a merged data set is probably too noisy to provide satisfactory classification results. In the second progress report of CODTRACE, large variation in parasite abundances between seasons was reported, but no consistent patterns accounting for such variation could be found. Further implementation of the method in a fishery management context would require steady sampling in designated areas and constant update of the predictive models.

All predictive models could assign cod from the Baltic Sea, Iceland and Norway to their respective harvest location efficiently (over 95% for the Baltic and Norwegian cod and about 90% for the Icelandic cod), but failed to deliver similar correct allocation rates in cod from the Celtic, Irish and North Seas. Unfortunately the datasets, except Autumn 2002, were unbalanced (some localities included more cod than others), being cod was from the Celtic and North Seas (Spring 2002 dataset) and Irish Sea (Spring 2003 dataset) underrepresented in the learning processes (Table 1). Since both the BBP network and LDA's ability to correctly classify cases depends on the number of examples for each category used in the training process, additional data from the Celtic, Irish and North Sea might probably improve the correct classification rates in the test sets. Unbalanced samples might also account for the apparently erratic behaviour of the correctly classified cod in the test set of cod from the Celtic Sea, ranging from as bad as 57% in the most unbalanced data set (Spring 2002) to 92% in the balanced Autumn 2002 dataset (Table 2). Note also the better overall performance of the BBP network with the balanced Autumn 2002 dataset that probably illustrate the full potential of the technique if sample sizes of each category are similar in model training.

Although the accuracy of the BBP network was similar to LDA, it is possible that the accuracy of the neural network could improve with larger sample sizes. Neural networks have traditionally been used to decipher and classify very subtle patterns in large datasets. The advantage being that the more "experienced" they become, the better they are at detecting patterns. This ability would be an asset in regular sampling programmes, greater resolution potentially arising out of their application to increasingly large amounts of data. However, room for prediction improvement is probably constrained by similarities in parasite assemblages of cod, particularly between the Irish, Celtic and North Sea. The fact that about 10% of the fish analysed was misallocated by two statistical methods based on quite different assumptions is a strong suggestion of an underlying unpredictability in the composition of parasite assemblages of cod that cannot solved by the sophistication of the statistical method applied. However, this report shows that even if this possibility is confirmed, analyses of parasite assemblages are useful to predict harvest location of Baltic, Icelandic or Norwegian cod and might still be crucial used in combination with other biomarker techniques developed under the CODTRACE project.

Contact

Ruiz Torres PEDRO, (Professor)
Tel.: +34-96-3864044
Fax: +34-96-3983021
E-mail
Follow us on: RSS Facebook Twitter YouTube Managed by the EU Publications Office Top