Classification of Autism Spectrum Disorder from rs-fMRI Scans by Feature Reduction and Ensemble Machine Learning Classifier

Autism Spectrum Disorder (ASD) is a multifaceted neurodevelopmental disorder and genetically heterogeneous. Neuroimaging techniques like rs-fMRI have shown changes in brain structure and connectivity in ASD patients. Regions associated with language processing, sensory perception, and social communication are frequently affected. Early diagnosis of ASD might prove vital in timely treatment of the patient. This study presents a novel approach for detecting ASD by employing Principal Component Analysis (PCA) and ensemble machine learning. The objective of this study is to mitigate the computational complexity associated with classifying individuals with ASD from typically developing (TD) individuals and increasing the accuracy of linear classifiers over existing studies. This study also highlights that the other ASD subtypes are distinct from Autism. Finding the optimum hyperparameters for four different supervised learning models – SVM, Linear SVC, Logistic Regression and Ridge classifier, an optimum ensemble learning model has been designed using the three most promising classifiers. Lastly, 10 fold cross validation has been performed to evaluate the five classifier models for the categorization of ASDTD and Autism-TD, with and without the inclusion of PCA. The ensemble model incorporating PCA, trained and tested on dataset excluding subtypes like Aspergers and PDD-NOS achieved the highest accuracy of 76.01%. The precision and recall of this model are 72.01% and 70.69% respectively. Overall, the results of this study demonstrate the potential for enhanced diagnostic precision of Autism with greater computational efficiency.