Modeling quality changes in heat-processed orange juice: a comparative study of artificial neural network and multiple linear regression approaches

The purpose of the study was to assess the prediction ability of multiple linear regression (MLR) and artificial neural network (ANN) models for the browning index, total carotenoid content, 5-hydroxymethylfurfural (HMF) and ascorbic acid of orange juice during storage after heat processing. For ANN models, the effect of neuron number of the hidden layer, epoch number, training algorithms and transfer functions are investigated using the trial-error method for selecting best design ANN models. The different methods (stepwise and enter) in multiple linear regression models were performed for detecting impact of independent variables on dependent variables. The performance of ANN and MLR models was determined through unseen data by means of statistical analysis. Regarding performance indices of ANN models for test data, overall R and R2 were recorded as follows: 0.92 and 0.84 (browning index), 0.99 and 0.98 (HMF), 0.92 and 0.86 (ascorbic acid), 0.97 and 0.94 (total carotenoid content), respectively. R and R2 values of MLR models for test data were 0.79 and 0.68 (browning index), 0.94 and 0.88 (HMF), 0.92 and 0.85 (ascorbic acid), 0.93 and 0.90 (total carotenoid content), respectively. Both models provided accurate predictions. However, the superior predictive power of ANN models is that they can learn directly from examples without calculating the parameters using statistical techniques. The results revealed that ANN models showed greater potential with high R and R2 value, and the lowest error values when compared with the MLR model, but both ANN and MLR had almost same performance for prediction of ascorbic acid.