Exploring the most discriminative brain structural abnormalities in ASD with multi-stage progressive feature refinement approach

Bing Sun, Yingying Xu, Siuching Kat, Anlan Sun, Tingni Yin, Liyang Zhao, Xing Su, Jialu Chen, Hui Wang, Xiaoyun Gong, Qinyi Liu, Gangqiang Han, Shuchen Peng, Xue Li, Jing LiuFrontiers in Psychiatry 2024 · DOI

Abstract

Structural irregularities in the reward system may play a crucial role in the pathophysiology of ASD, and specific hippocampal subfields may also contribute uniquely, warranting further investigation.

Introduction

Despite decades of neuroimaging research, a definitive "map" of brain structural changes in Autism Spectrum Disorder (ASD) remains elusive. The inherent heterogeneity of ASD, combined with the "high-dimensional" nature of MRI data—where the number of potential features (volume, thickness, surface area) often exceeds the number of participants—leads to inconsistent findings across studies. Furthermore, many investigations focus on the cerebral cortex, frequently neglecting subcortical structures like the hippocampus due to segmentation challenges. This study addresses these issues by employing a multi-stage progressive refinement approach to identify the most discriminative structural features, with a specific focus on detailed hippocampal subfields.

Methodology

The researchers utilized a dataset of 244 individuals (175 with ASD and 69 typically developing, aged 7–18 years) from Peking University Sixth Hospital. Brain images were processed using FreeSurfer v7.2.0 to segment 317 initial MRI features, including precise hippocampal subfields like the CA2 and CA3 regions.

To distill these features, a Multi-Stage Progressive Feature Refinement Approach was implemented:

  1. Filter Stage (SelectKBest): ANOVA F-values were used to identify the top 20% of features (64 features) most strongly correlated with ASD diagnosis.
  2. Embedded Stage (LASSO): The Least Absolute Shrinkage and Selection Operator (LASSO) addressed multicollinearity. Features consistently appearing across five-fold cross-validation were retained, reducing the pool to 27.
  3. Wrapper Stage (SVM-based Voting): A Support Vector Machine (SVM) model ranked the remaining features. Features were added progressively until the Area Under the Curve (AUC) stabilized.
LASSO coefficient paths illustrating the progressive feature selection process across different validation folds.

This layered approach balances statistical relevance, redundancy reduction, and predictive power more effectively than standard single-step methods.

Results and Key Insights

The final model identified nine discriminative features that effectively distinguished ASD from typically developing peers, achieving an AUC of 0.77 and an accuracy of 71.1% (Sensitivity: 68.0%, Specificity: 79.7%).

The Reward System and Social Motivation

A primary finding was the significant overlap between the identified features and the brain's reward circuitry. Key regions included the Nucleus Accumbens, Anterior Cingulate Cortex, and Temporal Pole. These structural differences support the "social motivation hypothesis," suggesting that irregularities in reward processing may drive the social communication challenges characteristic of ASD.

The Role of Hippocampal Subfields

The volume of the left CA2/3 hippocampal subfield emerged as a top discriminator. In animal models, the CA2 region is critical for social memory; its prominence here suggests a parallel biological mechanism in humans for recognizing and remembering others.

Subcortical vs. Cortical Metrics

While cortical thickness is a common focus in neuroimaging, this study found that subcortical volumes, such as the Nucleus Accumbens, provided more stable and discriminative power within the predictive model.

Discussion and Limitations

While the study provides a refined "signature" for ASD, several limitations remain. The sample was imbalanced, with a larger ASD group, and the broad age range (7–18 years) encompasses significant developmental changes. Additionally, the model lacks external validation on independent datasets. Future research should consider longitudinal designs and more complex deep learning architectures to further explore these nuanced structural relationships.

Summary updated

Abstract

Structural irregularities in the reward system may play a crucial role in the pathophysiology of ASD, and specific hippocampal subfields may also contribute uniquely, warranting further investigation.
BibTeX
@inproceedings{sun2024exploring,
  author = {Bing Sun and Yingying Xu and Siuching Kat and Anlan Sun and Tingni Yin and Liyang Zhao and Xing Su and Jialu Chen and Hui Wang and Xiaoyun Gong and Qinyi Liu and Gangqiang Han and Shuchen Peng and Xue Li and Jing Liu},
  title = {{Exploring the most discriminative brain structural abnormalities in ASD with multi-stage progressive feature refinement approach}},
  booktitle = {Frontiers in Psychiatry},
  year = {2024},
  doi = {10.3389/fpsyt.2024.1463654},
}
Figure