Multimodal Deep Learning Algorithms for Autism Prediction: Integrating Functional MRI and Gut Microbiome Data

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition that affects an individual's social interactions, communication abilities, and behavioral patterns. Recent studies suggest that both neurological and gastrointestinal factors contribute to ASD etiology, necessitating multimodal diagnostic approaches. However, integrating functional MRI (fMRI) and gut microbiome data remains a critical challenge, especially when datasets lack common individuals. This study explores different Deep Learning methods for multimodal ASD prediction, including Multimodal Variational Autoencoder (MVAE), Generalized Canonical Correlation Analysis (GCCA), and Graph Neural Networks (GNNs). Each algorithm addresses specific challenges: MVAE effectively handles missing modalities, GCCA aligns feature spaces across datasets without requiring paired samples, and GNNs leverage functional connectivity networks to extract graph-based neurobiological patterns. By applying these models to fMRI-derived connectivity matrices and microbiome composition profiles, we systematically evaluate their classification accuracy, robustness, and interpretability. Experimental results demonstrate that GNN-based approaches achieve the highest predictive accuracy (91.62%), outperforming conventional multimodal fusion techniques. MVAE provides strong latent representations but struggles with small, heterogeneous datasets, while GCCA offers a scalable framework for cross-modal correlation learning. This study emphasizes the potential of graph-based deep learning in diagnosing ASD and highlights the significance of integrating multimodal data in research on neurodevelopmental disorders. Future extensions may explore transformer-based architectures and domain adaptation techniques to enhance predictive performance further.

Topics & Concepts