Classification of Wild and Farmed Salmon Using Bayesian Belief Networks and Gas Chromatography-Derived Fatty Acid Distributions

In this study, we present the use of Bayesian Belief Networks (BBN) for the classification of wild versus farmed Atlantic salmon (Salmo salar L.). Using a data set of 131 salmon samples from several geographical origins and the gas chromatography-derived distributions of 12 fatty acids (FAs), a Bayesian Belief Network was constructed, ultimately using only the three most important FAs (16:1n-7, 18:2n-6, and 22:5n-3). The training data set yielded a prediction error of 0% (68/68 farmed; 20/20 wild correct) while the validation data set prediction error was 4.65% (32/32 farmed; 9/11 wild correct). Different randomly chosen validation sets yielded similar prediction accuracies. This model was then applied to 30 market (store-bought) samples where predictions were compared with the product labels.