Optimizing artificial neural network parameters using Taguchi method: An application of artificial heart pump

Abstract

Today, circulatory system diseases constitute a major problem in terms of people's quality of life. In patients diagnosed with heart failure, the primary treatment is heart transplantation, but artificial heart support pumps are also offered as a solution due to long waiting lists and a shortage of donors. In artificial heart support pumps, choosing the appropriate design is important for proper blood flow. In this study, the effects of pump rotational speed, blood flow rate, and pump blade angle parameters, which are used in the design of heart support pumps, on turbulence, shear stress, speed, and pressure values of the blood were investigated. Artificial neural network models were developed using analysis results obtained from ANSYS software, and the most successful artificial neural network models were determined by testing different numbers of neurons in the hidden layer, training functions, and training data rates. The modeling performance was evaluated based on the R² and mean squared error criteria, with the best results generally achieved using the Levenberg-Marquardt method. The highest modeling accuracy for turbulence, shear stress, velocity, and pressure was obtained with different neuron counts and training data rates; however, the precise effect of these parameters on model performance was not clearly defined. While the Levenberg-Marquardt method produced more consistent results compared to other methods, the scaled conjugate gradient and Bayesian regularization methods occasionally yielded successful outcomes but did not demonstrate stable performance.