Using Machine Learning to Predict Risk and Diagnostic Outcomes of Autism Spectrum Disorder

Background: Identifying early risk and/or diagnostic biomarkers for autism spectrum disorder is one of the challenges in autism research today. Recent studies have investigated the use of Event Related Potential ERP to identify neural markers of ASD in infancy, through identifying brain function difference between infants at risk for ASD and control. Previous findings indicated the existence of significant differences in statistical features and in power spectral density of the different frequency bands across all brain regions (frontal, central, parietal, occipital and temporal).

Objectives: In this study, we combined averaged ERP extracted measures, including power metrics as well as statistical features across different brain regions to explore the performance accuracy of classification algorithms in distinguishing on the one hand between infants At-risk with ASD and control group and secondly between infants diagnosed with ASD and others with no-ASD.

Methods: EEG data was collected at 6 months of age on 54 infants at risk for autism (HR) by virtue of having an older diagnosed sibling and 46 low risk (LR). From the group of HR infants, 17 were later diagnosed with ASD (n= 17 HR-ASD, 34 HR-noASD and 54 LR-noASD). ERP recordings were collected while infants watched a passive “eye-gaze” task. A total of 134 features, including power spectral density at different frequency bands (delta, theta, alpha, beta and gamma), P1, N1, P2, N270, P3, and LPC, were derived from the averaged ERP data during the averted eye gaze stimuli across all brain regions. Three classifiers including linear and nonlinear techniques: Discriminant analysis (DA), K-Nearest Neighbor (KNN) and Support Vector machines (SVM), are compared for distinguishing offline data to choose the suitable classification algorithm for each experiment. To avoid overfitting, 10-fold cross validation was used to test the efficiency of machine learning algorithms.

Results: the results indicate that while all methods succeeded in achieving suitable performance levels, DA 10-cross validation provides the best average accuracy rate for correctly associating an ERP infant to risk or control classes and SVM 10-cross validation shows the highest accuracy in distinguishing infants diagnosed with ASD from LR-noASD and HR-no ASD.

Conclusions: Due to the number of ERP components and conditions being integrated in this study, our analysis was limited due to the sample size. Optimizing these machine learning techniques to be applied at a large scale will allow for robust prediction of risk status and diagnostic outcomes based on a combination of ERP features.