PREDICTION OF ANTIOXIDANT ACTIVITY OF FRUITS THROUGH THE USE OF ARTIFICIAL NEURAL NETWORK

Phenolic compounds are recognized for their significant bioactivity, playing a role in plant protection against abiotic stresses and providing therapeutic benefits to human health. Although interest in functional foods has driven the demand for quality analysis, conventional analytical methods for determining antioxidant capacity are costly and require high operational time. In this context, the application of Artificial Intelligence tools emerges as a strategic alternative for process optimization. The present study aimed to predict the antioxidant activity of different fruit matrices using Artificial Neural Networks (ANNs). The database was structured from ten scientific articles, from which phenolic profiles and antioxidant activity values (DPPH) were extracted. For model development, a Multilayer Perceptron (MLP) architecture was used with 13 input variables and three hidden layers. Data were normalized using the Min-Max method and divided into training (70%) and testing (30%) sets. Different activation functions and optimization algorithms were tested to evaluate predictive accuracy. The results demonstrated that the combination of the hyperbolic tangent activation function (tanh) with the Stochastic Gradient Descent (SGD) optimizer yielded the best performance, achieving a Mean Absolute Percentage Error (MAPE) of 3.93×10⁻⁶ and a Mean Squared Error (MSE) of 2.28×10⁻¹⁰. It is concluded that ANNs constitute a robust, fast, and low-cost methodology for predicting antioxidant capacity, with high potential for application in quality control within the food industry.