[Paper Review] Functional connectivity patterns of autism spectrum disorder identified by deep feature learning

Hongyoon Choi|arXiv (Cornell University)|Jul 25, 2017

Functional Brain Connectivity Studies20 citations

TL;DR

This study proposes a deep learning-based data-driven approach using variational autoencoders (VAE) to extract multivariate, nonlinear functional connectivity patterns from resting-state fMRI data in autism spectrum disorder (ASD). It identifies a significant functional connectivity feature linked to frontoparietal, dorsal medial frontal cortex, and corticostriatal connections, which correlates with IQ and reveals previously undetected patterns of brain network disruption in ASD.

ABSTRACT

Autism spectrum disorder (ASD) is regarded as a brain disease with globally disrupted neuronal networks. Even though fMRI studies have revealed abnormal functional connectivity in ASD, they have not reached a consensus of the disrupted patterns. Here, a deep learning-based feature extraction method identifies multivariate and nonlinear functional connectivity patterns of ASD. Resting-state fMRI data of 972 subjects (465 ASD 507 normal controls) acquired from the Autism Brain Imaging Data Exchange were used. A functional connectivity matrix of each subject was generated using 90 predefined brain regions. As a data-driven feature extraction method without prior knowledge such as subjects diagnosis, variational autoencoder (VAE) summarized the functional connectivity matrix into 2 features. Those feature values of ASD patients were statistically compared with those of controls. A feature was significantly different between ASD and normal controls. The extracted features were visualized by VAE-based generator which can produce virtual functional connectivity matrices. The ASD-related feature was associated with frontoparietal connections, interconnections of the dorsal medial frontal cortex and corticostriatal connections. It also showed a trend of negative correlation with full-scale IQ. A data-driven feature extraction based on deep learning could identify complex patterns of functional connectivity of ASD. This approach will help discover complex patterns of abnormalities in brain connectivity in various brain disorders.

Motivation & Objective

• To identify complex, multivariate functional connectivity patterns in autism spectrum disorder (ASD) that are not captured by traditional linear methods.
• To overcome the lack of consensus in prior fMRI studies on ASD-related connectivity disruptions by using a data-driven, unsupervised deep learning approach.
• To extract low-dimensional, meaningful features from functional connectivity matrices without relying on prior diagnostic labels or predefined network models.
• To visualize and interpret the learned features using a VAE-based generator to produce synthetic functional connectivity matrices.
• To investigate the relationship between identified connectivity features and cognitive measures such as full-scale IQ in ASD.

Proposed method

• Functional connectivity matrices were constructed from resting-state fMRI data using 90 predefined brain regions for each of 972 subjects (465 ASD, 507 controls) from the ABIDE dataset.
• A variational autoencoder (VAE) was applied to each subject's connectivity matrix to learn a low-dimensional, continuous latent representation of 2 features, capturing nonlinear and multivariate patterns.
• The VAE was trained in an unsupervised manner, without using diagnostic labels, to preserve the intrinsic structure of functional connectivity data.
• The latent features were statistically compared between ASD and control groups to identify significant differences.
• A VAE-based generator was used to visualize the learned features by generating synthetic functional connectivity matrices corresponding to extreme values of the identified ASD-related feature.
• The relationship between the ASD-related feature and full-scale IQ was assessed via correlation analysis.

Experimental results

Research questions

• RQ1 What multivariate, nonlinear functional connectivity patterns in ASD can be identified using a data-driven deep learning approach?
• RQ2 Which brain network connections are most strongly associated with the identified functional connectivity feature in ASD?
• RQ3 How does the identified functional connectivity feature relate to cognitive performance, particularly full-scale IQ?
• RQ4 Can a deep generative model effectively visualize and interpret the functional connectivity patterns learned from ASD data?
• RQ5 To what extent does the VAE-based feature extraction method reveal consistent and biologically plausible connectivity disruptions in ASD beyond traditional univariate analyses?

Key findings

• A single functional connectivity feature extracted via VAE showed a statistically significant difference between ASD patients and neurotypical controls (p < 0.05).
• The ASD-related feature was most strongly associated with frontoparietal connections, interconnections of the dorsal medial frontal cortex, and corticostriatal pathways.
• The identified feature exhibited a negative trend in correlation with full-scale IQ, suggesting a potential link between disrupted connectivity and cognitive ability in ASD.
• The VAE-based generator successfully produced realistic virtual functional connectivity matrices that reflected the ASD-related pattern, enabling visual interpretation of the learned feature.
• The data-driven deep learning approach uncovered complex, nonlinear connectivity patterns in ASD that were not consistently detected by conventional fMRI connectivity analyses.
• The method demonstrated the potential to reveal subtle, multivariate brain network abnormalities in neurodevelopmental disorders without relying on prior assumptions about network structure.