Detecting Autism Spectrum Disorder from Functional Magnetic Resonance Imaging Data with an Ensemble Machine Learning Method

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder caused by functional impairment of the brain. The prevalence of this disease in the population has been increasing in recent years. Early identification of ASD can help children improve with early behavioral interventions. However, traditional testing methods are very expensive and time-consuming. Machine learning methods based on functional Magnetic Resonance Imaging(fMRI) have emerged as a research direction for autism-assisted diagnosis. In this paper, we developed an ensemble approach to distinguish ASD from normal controls by a stacking strategy, integrating the results of multiple machine learning classifiers. The dataset included 569 subjects from different centers in the study. We extracted the functional connectivity (FC) as features and removed the redundant ones by SVM-RFE, building an ensemble model for prediction by the stacking strategy. Specifically, KNN, NB, RF, GBDT, and XGBoost were used as base models, and logistic regression was employed as a meta-model. The experimental results illustrated that the ensemble method could improve the evaluation metrics of prediction results by about 2% on average compared with individual classifiers. Thus, the stacking strategy has better classification performance for intelligent diagnosis of ASD via fMRI.