Performance comparison of artificial neural networks learning algorithms and activation functions in predicting severity of autism

Abstract

Artificial neural networks are one of the most efficient methods for pattern recognition and have a vast range of applications for aiding medical decision making. The proposed study applies feed-forward back-propagation neural networks as a classifier and compares the combination of nine learning algorithms and three activation functions to build a knowledge-based system with the best network architecture for predicting the severity of autism. The performances of the derived models were evaluated based on statistical criteria such as mean squared error (MSE), mean absolute percentage error (MAPE), root mean squared error (RMSE), regression (R value), training time and number of epochs. The study findings showed that the optimal performance was achieved by model MLP_LM_104 trained on Levenberg–Marquardt (LM) back-propagation algorithm having network topology of 40-10-4 with purelin and tansig activation functions in hidden and output layers. The regression coefficients for training, validation and test datasets were 0.996, 0.996 and 0.994, respectively. The MSE, RMSE and MAPE were \( 2.26 \times 10^{ - 4} \), \( 1.50 \times 10^{ - 2} \) and \( 1.13 \), respectively. Furthermore, BFGS quasi-Newton (BFG), conjugate gradient, gradient descent and resilient back-propagation (RP) algorithms did not perform well. Models trained with BFG algorithms required longer training time, whereas the performance of models trained on RP algorithm got worse as the numbers of hidden neurons were increased.