Structural connectome differences between autism and neurotypical control groups using autoencoder

359 

Abstract Submission 

Yurim Jang¹, Hyoungshin Choi^2,3, Seulki Yoo⁴, Hyunjin Park^3,5, Bo-yong Park^6,7,3

¹Artificial Intelligence Convergence Research Center, Inha University, Incheon, Korea, Republic of, ²Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, Republic of Korea, ³Center for Neuroscience Imaging Research, Institute for Basic Science, Suwon, Korea, Republic of, ⁴Convergence Research Institute, Sungkyunkwan University, Suwon, Republic of Korea, ⁵School of Electronic and Electrical Engineering, Sungkyunkwan University, Suwon, Korea, Republic of, ⁶Department of Data Science, Inha University, Incheon, Korea, Republic of, ⁷Department of Statistics and Data Science, Inha University, Incheon, Korea, Republic of

Yurim Jang  
Artificial Intelligence Convergence Research Center, Inha University
Incheon, Korea, Republic of

Hyoungshin Choi  
Department of Electrical and Computer Engineering, Sungkyunkwan University|Center for Neuroscience Imaging Research, Institute for Basic Science
Suwon, Republic of Korea|Suwon, Korea, Republic of

Seulki Yoo  
Convergence Research Institute, Sungkyunkwan University
Suwon, Republic of Korea

Hyunjin Park  
Center for Neuroscience Imaging Research, Institute for Basic Science|School of Electronic and Electrical Engineering, Sungkyunkwan University
Suwon, Korea, Republic of|Suwon, Korea, Republic of

Bo-yong Park  
Department of Data Science, Inha University|Department of Statistics and Data Science, Inha University|Center for Neuroscience Imaging Research, Institute for Basic Science
Incheon, Korea, Republic of|Incheon, Korea, Republic of|Suwon, Korea, Republic of

Autism spectrum disorder is a pervasive condition during development. Individuals with autism show deficits in sensory and social communication skills [1]. Recent neuroimaging studies based on magnetic resonance imaging (MRI) found alterations in large-scale functional brain networks using low-dimensional features [2], [3]. Here, we aimed to assess network disorganization of the brain in individuals with autism by generating low-dimensional latent features of structural connectivity using an autoencoder.

We obtained diffusion MRI of 80 individuals with autism (mean ± standard deviation (SD) age = 12.1 ± 4.9 years) and 61 neurotypical controls (13.2 ± 4.0 years) from the Autism Brain Imaging Data Exchange-II (ABIDE-II) initiative [4]. The diffusion MRI was preprocessed using MRtrix3 [5], and the structural connectivity matrix was constructed based on probabilistic tractography with 200 brain regions defined using the Schaefer atlas. After controlling for age, sex, and site from the structural connectivity, we trained the autoencoder, which consisted of the encoder and decoder layers. The encoder reduces high-dimensional input data to generate low-dimensional latent features, and the decoder reconstructs the original data using the latent features. The model consisted of five encoder and decoder layers, and we used Averaged Stochastic Gradient Descent (ASGD) optimizer with a learning rate of 0.00008. The model was trained for the autism and neurotypical control groups, respectively. The performance of the model was assessed based on the linear correlations between the original and reconstructed data. We then calculated the integrated gradient values, which denote the attribution of each element of the input connectivity matrix for predicting the latent features [6]. After the z-normalization, we compared the integrated gradient values between autism and control groups using two-sample t-tests and 1,000 permutation tests. The multiple comparisons were corrected using a false discovery rate (FDR) < 0.05. We adopted canonical correlation analysis (CCA) to investigate the association between the integrated gradient values and the symptom severity of autism measured by the Autism Diagnostic Observation Schedule (ADOS). The optimal number of canonical components was determined using five-fold cross-validation, and the degrees of the associations were assessed using the explained variance.

The autoencoder model revealed significant correlations between the original and reconstructed data for autism (mean ± SD; 100 bootstraps = 0.427 ± 0.165) and control groups (0.271 ± 0.133) (Fig.1A). Between-group comparison of the integrated gradient values showed the highest effects within the default-mode network and between the visual and frontoparietal/ventral networks, while the smallest effects were found within the visual network and between the somatomotor and dorsal attention networks (Fig.1B). The brain-behavior associations revealed that the three canonical components were significantly associated (1st: r = 0.726, p < 0.001; 2nd: r = 0.732, p < 0.001; 3rd: r = 0.647, p < 0.001; Fig. 2). In particular, the ADOS communication sub-score was strongly associated with the integrated gradient values within the default-mode network and between somatomotor-visual/limbic/frontoparietal networks.

We identified structural connectivity differences between individuals with autism and neurotypical control via low-dimensional latent features of the autoencoder. In particular, sensory and transmodal regions showed significant between-group differences. Additionally, the differences were related to the communication skills in autism. Our findings may improve the understanding of autism.
Funding:
NRF-2021R1F1A1052303; NRF-2022R1A5A7033499, IITP funded by (MSIT) (No. 2022-0-00448, ; No. RS-2022-00155915, (Inha University); No. 2021-0-02068, Artificial Intelligence Innovation Hub), IBS-R015-D1.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ¹

Connectivity (eg. functional, effective, structural) ²

Diffusion MRI

Autism

Machine Learning

STRUCTURAL MRI