Modeling of reversed-phase direct immersion microextraction of acrylamide by feed forward neural network

Acrylamide even in trace amounts is a potential cause of cancer in humans and can be created in many foodstuff that are cooked at high temperatures [1], therefore it is of importance to detect and quantify it in brown bread and potato chips and crisps. In this work, a novel reversed-phase direct immersion single drop microextraction was developed, optimized and used for the determination of acrylamide at low levels in potato crisps samples [2]. Then, a computational modeling method based on a feed-forward neural network was established to construct a predicting model for the determination of acrylamide using the results obtained empirically. The inputs of feed-forward neural network model were pH, extraction time, extraction temperature, stirring rate, drop volume, and sample volume; and the output was peak areas of acrylamide in the chromatograms. The BFGS Quasi-Newton algorithm was used to train feed forward neural network by the patterns gathered through experiments. The patterns used for modeling were divided in three subsets: 70% for training data set, 15% for validation data set, and 15% for testing data set. As an activation function, hidden neurons use hyperbolic tangent sigmoid function tansig(s)= 2 1 1 2    s e , where   1 tansig(s) 1  and output neuron uses a linear transfer function purelin(s)=s where  purelin(s)   . A neural network with 4 hidden neurons was considered for modeling purpose. In order to assess the efficiency of model for the prediction of acrylamide, root mean square error (RMSE) and determination coefficient (R2) were used. The RMSE of obtained predictor model in training data was 32.0, in validation data set was 34.15, and in testing (unseen) data set was 36.35. A regression plot is created for each of three mentioned data sets to verify the relationship between forecasted outputs by model and actual outputs of patterns. The resulted R2 value was 0.92, 0.90, and 0.90 for training, validation and unseen data sets, correspondingly that means there is a linear relation between the outputs of network and experimental outputs.