Functional Connectivity based Machine Learning Approach for Autism Detection in Young Children using MEG Signals

Objective: Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder, and identifying early autism biomarkers plays a vital role in improving detection and subsequent life outcomes. This study aims to reveal hidden biomarkers in the patterns of functional brain connectivity as recorded by the neuro-magnetic brain responses in children with ASD.

Approach: We recorded resting-state MEG signals from thirty children with ASD (4-7 years) and thirty age, gender-matched typically developing (TD) children. We used a complex coherency-based functional connectivity analysis to understand the interactions between different brain regions of the neural system. The work characterizes the large-scale neural activity at different brain oscillations using functional connectivity analysis and assesses the classification performance of coherence-based (COH) measures for autism detection in young children. A comparative study has also been carried out on COH-based connectivity networks both region-wise and sensor-wise to understand frequency-band-specific connectivity patterns and their connections with autism symptomatology. We used Artificial Neural Network (ANN) and Support Vector Machine (SVM) classifiers in the machine learning framework with a 5-fold cross-validation technique.

Main results: To classify ASD from TD children, the COH connectivity feature yields the highest classification accuracy of 91.66% in the high gamma (50-100 Hz) frequency band. In region-wise connectivity analysis, the second highest performance is in the delta band (1-4 Hz) after the gamma band. Combining the delta and gamma band features, we achieved a classification accuracy of 95.03% and 93.33% in the ANN and SVM classifiers, respectively. Using classification performance metrics and further statistical analysis, we show that ASD children demonstrate significant hyperconnectivity.

Significance: Our findings support the weak central coherency theory in autism detections. Further, despite its lower complexity, we show that region-wise coherence analysis outperforms the sensor-wise connectivity analysis. Altogether, these results demonstrate the functional brain connectivity patterns as an appropriate biomarker of autism in young children.