Identifying autism spectrum disorder based on machine learning for multi-site fMRI

Abstract

Background

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by repetitive stereotypical behavior and social impairment. Early diagnosis is essential for developing a treatment plan for autism. Although multi-site data can expand the dataset to facilitate the process of data analysis, data heterogeneity between sites and the large amount of data make data analysis difficult.

Method

To address these issues, this paper proposes a multi-site autism identification method based on machine learning technique. Firstly, the fMRI data from all sites are converted into a glass brain dataset and their features are extracted with LeNet5. Then, the extracted glass brain features are used to construct a partial correlation matrix at subject-level and the multi-site dataset are constructed by feature selection, which is finally classified using MLP. In order to alleviate the heterogeneity of the data and improve the accuracy of data classification, a new dataset partitioning method, Split-Merge-Split (SMS), is proposed in this paper to reduce the variability between the features extracted by the model in the training and test sets.

Results

Extensive quantitative and qualitative evaluations demonstrate the proposed method enhanced the recognition accuracy on both single-site and multi-site dataset, which shows the effectiveness of this method. Specifically, in single-site classification, our method achieved its highest accuracy at the OHSU site, reaching an accuracy of 93 %. In multi-site classification, our method attained an accuracy of 83.5 %.