Research On Prediction Of Autism Spectrum Disorder Based On Multi-Task Learning

Autism spectrum disorder(ASD)is a complex neurodevelopmental disorder.To date,there is still no objective and unified diagnostic method.The release of multi-center magnetic resonance imaging datasets on ASD and the development of machine learning techniques have greatly promoted the objective diagnosis of ASD.However,due to the heterogeneity between multi-center data resulted from inconsistent collection standards and the heterogeneity of the disease itself caused by age and IQ,the objective diagnosis of ASD has not yet achieved satisfactory results.Based on multi-center multi-modal data of ASD,this thesis first regards the prediction of each center as a learning task,and proposes two multi-task learning methods with the purpose of reducing the influence of heterogeneity among multi-center data.Then in order to further reduce the effect of the disease heterogeneity,a method combining multi-source domain adaptation and multi-task learning is proposed,and uses the prediction of each age and IQ group as a learning task.The main work is as follows:(1)This thesis proposes a multi-task learning method that introduces the relationship between features to diagnose ASD and predict its severity.Existing ASD multi-task learning methods only considered the relationship between tasks and the relationship between modalities,but did not utilize the constraints within the data.To this end,this thesis explores the similarity constraints between features to use the deep information of the data for improving the prediction performance.The proposed method surpasses other methods in performance.The average accuracy of ASD diagnosis is 70.99%,and the average correlation value of ASD severity prediction is 0.55.(2)The existing methods only focused on the shared features of multiple centers and ignored the features of each center.This thesis proposes a multi-task learning method based on shared features and center-specific features to obtain more accurate discriminant coefficients to improve the diagnostic performance of ASD.The discriminant coefficient matrix is decomposed into a shared feature matrix and a center-specific matrix.Row-sparse constraints are applied to the shared feature matrix to learn multi-center shared features.Element-sparse constraints are used to the center-specific matrix to learn center-specific features.In addition,a low rank constraint is further applied to the shared feature matrix of multi-modal data to capture the relationship between multiple related tasks.The proposed method achieves an average diagnostic accuracy of 71.84%,which is significantly better than other methods.(3)In order to jointly reduce the impacts of heterogeneity between ASD multi-center data and heterogeneity of ASD,this thesis proposes an ASD prediction method that combines multi-source domain adaptation and multi-task learning.First,a single-source domain adaptation method is improved to a multi-source domain adaptation method that projects the multi-center data into the same space to reduce the distribution difference between them.Subjects are then grouped according to age and verbal IQ,and the multi-task learning method proposed above is used to learn group-shared features and group-specific features for ASD diagnosis of each group.Based on the data of all age groups from the four centers,the proposed method reaches an average diagnostic accuracy of 75.13%,and has better performance compared to other comparison methods.The three proposed methods improve the predictive performance of ASD,and can help better clinicians in clinical diagnosis,thus promoting the understanding of the pathological mechanism of ASD.