Computational modeling of liquid phase microextraction of malondialdehyde from human blood plasma by multi-layer perceptron

Lipid peroxidation is the oxidative degradation of lipids. It is the process in which free radicals attack the lipids in cell membranes, resulting in cell damage. Malondialdehyde (MDA) is a compound which is released in human plasma during this process and can be used as an indicator of the level of lipid peroxidation. Therefore, determination of MDA is of high importance [1]. In this study, first MDA extraction from blood plasma was performed by the salt saturated single drop microextraction (SS-SDME) and its concentration was quantitatively determined by gas chromatography. All parameters of SS-SDME were optimized [2]. The data obtained here, were used to construct a predictive model based on multi-layer perceptron (MLP) for the determination of MDA in blood plasma. The inputs of MLP were stirring rate, pH, micro-drop volume, the volume of sample solution, extraction time, and temperature; and the output was relative peak area of MDA in chromatogram. The set of patterns used for modeling were 33 experiments and were divided in three subsets: 70% as training, 15% as validation, and 15% as testing sets. Given the training set, the scaled conjugate gradient method was used to train the MLP. The training epochs continue until the root mean square error (RMSE) between outputs of MLP and targets did not decrease during 6 continues epochs for patterns in validation set. As an activation function, hidden neurons use the 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)  . By performing some experiments, the number of 5 hidden neurons was considered for the MLP. The obtained RMSE for training, validation, and testing sets were 8.652259, 21.432824, and 16.158811, correspondingly. Regression diagrams were plotted for the three mentioned sets to investigate the relationship between predicted outputs of MLP and targets. The values of determination coefficients were 0.976524, 0.905604 and 0.905604 for training, validation and unseen data sets that states a good linear relation between outputs and targets. The linear range of the calibration curve was between 10-1000 μg.L-1, with a detection limit of 0.76 μg.L-1. At optimized conditions, a relative standard deviation of 8.66% was calculated for 5 successive analysis. The resulted predictive MLP model showed that it can save time, energy, and laboratory costs for optimization of the SS-SDME extraction of MDA from human blood plasma.