Classification of ASD Based on fMRI Data Using Two Hybrid Learning Approaches

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition marked by challenges in communication, heightened sensitivity to sensory input, difficulties in adapting to change, and atypical patterns of play. Since traditional methods of diagnosing this disorder rely heavily on behavioral assessments, they are highly prone to errors. Consequently, it is necessary to provide a reliable method for diagnosing this mental disorder by leveraging advanced machine-learning techniques. In this paper, two hybrid learning approaches that only use fMRI data have been designed and implemented. The first method is the simultaneous training of a deep autoencoder and a single-layer perceptron (SLP) to extract beneficial features and properly tune the model's hyperparameters for classification. The second method combines a support vector machine model with a data augmentation strategy using simultaneous training of a generative adversarial network and a deep autoencoder to prevent model overfitting. It is worth mentioning that in the two proposed methods, the Extra-Trees algorithm is initially used for dimensionality reduction and feature selection. These two methods have been implemented on the entire publicly available dataset released by the Autism Brain Imaging Data Exchange, and both results have been compared. Our study in this paper achieved a maximum accuracy of 70.8%, sensitivity of 67.4%, and specificity of 75.3%. These results indicate that our proposed approaches have performed significantly better than many recent studies in this field.