Autoencoder Based Methods for Diagnosis of Autism Spectrum Disorder

Abstract

Autism Spectrum Disorder (ASD) is a neurological disorder that affects a person’s behavior and social interaction. Integrating machine learning algorithms with neuroimages a diagnosis method can be established to detect ASD subjects from typical control (TC) subjects. In this study, we develop autoencoder based ASD diagnosis methods. Firstly, we design an autoencoder to extract high-level features from raw features, which are defined based on eigenvalues and centralities of functional brain networks constructed with the entire Autism Brain Imaging Data Exchange 1 (ABIDE 1) dataset. Secondly, we use these high-level features to train several traditional machine learning methods (SVM, KNN, and subspace discriminant), which achieve the classification accuracy of 72.6% and the area under the receiving operating characteristic curve (AUC) of 79.0%. We also use these high-level features to train a deep neural network (DNN) which achieves the classification accuracy of 76.2% and the AUC of 79.7%. Thirdly, we combine the pre-trained autoencoder with the DNN to train it, which achieves the classification accuracy of 79.2%, and the AUC of 82.4%. Finally, we also train SVM, KNN, and subspace discriminant with the features extracted from the combination of the pre-trained autoencoder and the DNN which achieves the classification accuracy of 74.6% and the AUC of 78.7%. These results show that our proposed methods for diagnosis of ASD outperform state-of-the-art studies.