Early-life autism spectrum disorder diagnosis with nonlinear EEG analysis using machine learning methods
With the increased prevalence of autism spectrum disorder (ASD) comes the challenge of identifying children at risk for ASD as early in life as possible so that they can benefit from early intervention. Currently, diagnosis is determined by measuring behaviors that often do not emerge until toddler or preschool age, but emerging research using neuroimaging suggests that brain changes occur much earlier. My senior thesis is focused on using high-density, task-related electroencephalography (EEG) data collected from 12-month-old infants to detect future ASD using neural activity instead of behavioral symptoms with the hope that earlier identification will lead to more effective treatment. Electrical activity recorded from the brain has complex dynamic properties that traditional linear analyses are not able to quantify. Therefore, this study uses nonlinear measures, including entropy and fractal dimension, computed from preprocessed EEG signal as features in a machine learning algorithm aimed to differentiate ASD from non-ASD outcomes. Previously, Bosl and colleagues (2018) had success using this method to classify ASD outcomes with baseline, non-task-related EEG data. However, a language task may be more sensitive to outcome prediction accuracy than baseline EEG because language is a domain frequently affected in ASD. A cross-validated support vector machine predicted ASD diagnosis of high-risk infants (those with an older sibling with ASD) with 95.5% accuracy. Sensitivity, specificity, PPV, and NPV rates were all over 92%. These results suggest that early brain function may be indicative of later ASD diagnosis, demonstrating potential for early risk assessment before observable behaviors of ASD emerge.