EEG Features Learned by Convolutional Neural Networks Reflect Alterations of Social Stimuli Processing in Autism

The P300 response exhibits abnormalities in autism spectrum disorder (ASD) and can potentially be used as ASD biomarker. Machine learning approaches represent useful tools for processing EEG signals by associating a cognitive state to the input neural activity. Interpretable convolutional neural networks (ICNNs) are growing interest among neuroscientists as they automatically learn the most relevant EEG features directly from the neural time series, and the learned features belong to easily interpretable domains (e.g., frequency domain). By leveraging on the ICNN features, data-driven analyses can be conducted, potentially helping in the characterization of novel biomarkers. However, ICNNs are mainly validated on healthy participants or on patients without relating the learned features with clinical scores (e.g., ADOS score in ASD). In this study, we analyzed the features learned by an ICNN trained to detect P300 from EEG recordings of 15 ASD participants. Interpretable spectral and spatial features were extracted and used to define ICNN-derived measures. The ICNN-derived spatial measure at Pz, but not spectral measures, was found to be positively correlated to ADOS scores. Moreover, a right-lateralization was found in the spatial measures, and resulted negatively correlated to ADOS scores, suggesting that the right-lateralization, reflecting the neural processing of complex social stimuli, is progressively disrupted as the severity of ASD increases. Our results suggest that features learned by ICNNs are sensitive to alterations in event-related potentials, and that could be exploited in future studies as potential candidates for conceiving novel ASD biomarkers.