Beyond the Gut Feeling: Machine-Learning Methods for Autism Screening

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that manifests itself as speech impairment and difficulties in social and cognitive skills. The onset of ASD is typically before the age of three, but many children are not diagnosed until later in life due to the current diagnosis being based on behavioral assessments. However, early intervention is crucial for minimizing the impact of ASD on both the child and their caregivers, and for developing an effective treatment plan for symptoms. We propose a gut microbiome-based screening tool for ASD utilizing feature selection, feature engineering, and machine learning algorithms. Our approach significantly reduces the feature space to 1-2% of its original size, emphasizing computational efficiency and informative attribute focus. Training and testing on two diverse datasets revealed minimal common features, underscoring the impact of regional differences on taxonomic behavior. Biomarker analysis confirmed known associations and unveiled novel connections, contributing valuable insights to the ongoing debate on ASD biomarkers. Notably, the genus Oscillospira has been identified by our Attribute Selection tools as being increased in ASD. Model performance, measured by Area Under the Curve, showcases the Bayes-Net algorithm on the Hierarchical Feature Engineering subset as the most optimal for ASD prediction. Our findings surpass existing literature, demonstrating the effectiveness of our approach in decoding the complex interplay between gut microbiome and ASD.