An Autism Spectrum Disorder Prediction Framework Based On Recurrent Neural Network

In recent years,the prevalence of autism spectrum disorders shows an upward trend year by year.Researchers combined with deep learning methods have achieved good results in the prediction of autism spectrum disorders.In this work,we focus on the prediction of autism spectrum disorders,study the processing of unbalanced eye movement data and three different prediction frameworks of autism spectrum disorder.The main work is completed as follows:(1)Based on Synthetic Minority Oversampling Technique(SMOTE)and Borderline-SMOTE,a synthetic oversampling technique is proposed to increase and distinguish the sample data.Based on the idea of cost-sensitive,the cost function of the model is changed,and the penalty factor is given to the data that is differentiated by the oversampling technique,which makes the model suitable for the unbalanced data.(2)A simple,highly modular framework for predicting autism spectrum disorders is proposed based on a dual-path network.Using the residual path and the dense connection path,the features of the eye movement data are reused and the new features are explored.Improve the classification performance by replacing the activation function,adjusting the order of the activation layers and the batch normalization layers.(3)Based on simple recurrent unit(SRU)and high order recurrent neural networks,a high order simple recurrent neural network is proposed.Based on the idea of simple recurrent unit,we reduce the dependence of hidden states and increase the parallelism of the network.Based on higher order recurrent neural networks,more memory cells are added to increase the feature extraction capability of the network and improve the accuracy of model classification.(4)Based on the improved dual path network and higher order simple recurrent neural network,a reliable and modular recurrent neural network for prediction of autism spectrum disorder is proposed.The dual path network is used to extract the coordinate correlation of eye movement data,the temporal order of the eye movement data is extracted by the higher order simple recurrent neural network,and the ability of the network to generate new features is enhanced by the model fusion.Experimental results show that in the dual-path network experiment,the improved cost function increases the AUC of the model by 11.79%.Replacing ReLU with SELU improves the classification accuracy from 87.97%to 89.49%.Adjusting the order of the activation layers and the batch normalization layers makes the model classification accuracy further improved to 90.88%.Compared with LSTM network,the training time of higher order simple recurrent neural network is shortened by 29.94%and the classification accuracy is improved from 88.68%to 92.47%.The classification accuracy of the framework of autism spectrum disorder based on recurrent neural network is as high as 93.65%.