A Novel Grading System for Autism Severity Level Using Task-Based Functional MRI: A Response to Speech Study

• Published in: IEEE access Vol. 9; pp. 100570 - 100582
• Main Authors: Haweel, Reem, Shalaby, Ahmed M., Mahmoud, Ali H., Ghazal, Mohammed, Seada, Noha, Ghoniemy, Said, Casanova, Manuel, Barnes, Gregory N., El-Baz, Ayman
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Activation analysis; Autism; Brain; Feature extraction; FSL; Functional magnetic resonance imaging; GLM; Indexing; Machine learning; Magnetic resonance imaging; Oversampling; Pediatrics; Schedules; Sensitivity; Speech; Task analysis; Variable speed drives
• DOI: 10.1109/ACCESS.2021.3097606
• ISSN: 2169-3536

Autism spectrum disorder (ASD) is a neuro-developmental disorder associated with impairments in social and lingual abilities. Failure in language development is variable in the ASD population and follows a wide spectrum. The autism diagnostic observation schedule (ADOS) is the current gold standard for diagnosing plus expert clinical judgment. Currently, studies aim to develop objective computer-aided technologies to diagnose autism with brain imaging modalities and machine learning. Task-based fMRI is a discriminating image modality that measures the functional activation of the brain. Computer-aided diagnosis systems aim to classify autistic subjects against typically developed peers despite the fact that autism is defined over a wide spectrum. Here, we propose a novel computer-aided grading framework in infants and toddlers (between 12 and 40 months) dependent on the analysis of brain activation in response to a speech experiment. First, brain activation responses are analyzed for 157 autistic subjects divided into three groups of: 92 mild, 32 moderate,and 33 severe as defined by ADOS calibrated severity score. Increased hypoactivation of the superior temporal cortex, angular gyrus, primary auditory cortex and cingulate gyri is exhibited with increasing autism spectrum severity. Less lateralization is also present when activation of the left hemisphere regions is recorded. Second, only these region of interest (ROI) areas are included for further local and global feature extraction in our ASD grading system. A comprehensive, two-stage system is developed using different classifiers. Four-fold cross-validation is adopted for testing. The first stage discriminates between moderate and the other two groups with an accuracy of 0.83 (sensitivity = 0.73, specificity = 0.83). Subsequently, a second stage classifies subjects as mild or severe autism with an accuracy of 0.81 (sensitivity = 0.81, specificity = 0.76). Finally, two validation techniques of synthesizing for oversampling and e of multiple random training and testing sets were adopted. The validation results proved the robustness of the proposed framework for an early computer-aided grading system to place subjects on the autism spectrum.