Prediction Of Autism-severity Based On Stacked Sparse Autoencoder

Autism,also known as Autism Spectrum Disorder(ASD),has many symptoms.Its main symptoms are social communication disorders,language communication disorders,narrow interests,and repetitive/stereotyped behaviors.Autism not only causes serious harm to the child,but also brings great psychological pressure and economic burden to the family.Current research indicates that early intervention has great value for autism,but the cause of autism is not clear,so early prediction and auxiliary diagnosis have become important research directions.At present,inconsistent diagnostic procedures and parental subjective information may affect doctors and lead to unreliable diagnostic results.We used the influencing factors of autism to predict the severity and the data came from the Shanghai Mental Health Center.We established a relationship model between the factors affecting the autism and the severity of autism,that is,selecting the relevant autism influencing factors as input variables,and using the grade of the Childhood Autism Rating Scale(CARS)as the output variable.In order to reduce the impact of low data quality on model training,we proposed an noise sample detection algorithm(OMND)for ordered multi-classification data based on the orderly characteristics of autism data.This algorithm takes into account the fact that the noise samples of different severity have different effects on the model training,and combines the distance weights to calculate an abnormal score for each sample of different classes.Samples with higher scores are then treated as noise samples and deleted,thereby minimizing the costly mispredictions of mispredicting high severity samples to low severity samples.Using the high-quality sample set after removing noise samples,we further constructed a deep neural network model(SSAE)of autism influencing factors and autism severity based on stacked sparse autoencoder and Softmax classifier.First,we preprocessed the input data,and then built and train several sparse autoencoders sequentially,that is,the hidden layer output of each sparse autoencoder is used as the input of the next sparse autoencoder to obtain the effective high-order features of the input data.Then we input the trained high-order features into the Softmax classifier,and used the label of severity for supervised learning.Next,the constructed deep neural network model is fine-tuned to optimize the network parameters.Finally,we used the grid search method and cross validation method to optimize the hyperparameters such as the number of hidden layers and hidden layer units,further determined the network structure and improved the accuracy of the prediction model of autism severity.Based on the constructed predictive model,the method of experimental contrast is used to discuss and analyze the prediction.First,we prove the effectiveness of the proposed OMND algorithm for the prediction of autism severity.We obtained three different sample sets and then used them to train the model separately.We compared the results of these three models and proved that the proposed OMND algorithm can make the model obtain higher prediction accuracy.Then,the effects of various learning algorithms,regular term coefficient,sparsity constraint degree and Dropout probability on the model prediction are analyzed.Finally,the proposed model was compared with the decision tree and the support vector machine multiclassification model,which proved that the proposed model has a higher classification accuracy.