Self-Supervised Machine Learning Approach for Autism Detection in Young Children using MEG Signal

Objective: Autism spectrum disorder (ASD) encompasses a broad spectrum of developmental disabilities and is associated with aberrant anatomical and functional neural activity patterns. Early detection of autism in young children is crucial for managing its impact effectively, but obtaining brain data from ASD children presents significant challenges. This study aims to develop an automatic and non-invasive method for detecting autism using magnetoencephalogram (MEG) signals from 30 children (4-7 years) with ASD and 30 age-matched typically developing (TD) children.

Methods: We employed a self-supervised learning (SSL) based machine learning framework, which has been successfully used in diverse domains. Specifically, we utilized a cross-domain (speech signal) pre-trained SSL architecture (wav2vec 2.0), leveraging the outputs of its final transformer layer as an embedding extractor for features in our classification. These SSL embedding (SSLE) features were then applied to artificial neural networks and support vector machine classifiers to distinguish between ASD and TD children.

Results: Our results show that the SSLE features yielded an overall classification accuracy of 88.39% attained by neural networks, exceeding the performance of commonly used hand-crafted features extracted from broadband brain oscillations.

Conclusion: The results show that even without in-domain pre-training, SSLE representations effectively discriminate between ASD and TD classes in this experiment. Pretrained SSL architectures, requiring minimal labeled data, outperformed conventional supervised models trained with much larger labelled data. Therefore, the proposed novel approach based on SSLE features shows promise for future autism detection research, especially in scenarios where collection of large data sets may be difficult.