| Title:
|
Dynamic Effective Connectivity for Autism Diagnosis and Severity Prediction via a Spatial-Temporal Graph Transformer
|
| Author(s):
|
1
|
| Subject:
|
Autism and Brain Mapping
|
|
|
|
Presenter Author:
|
|
| Name:
|
ستاره رضاخانی
|
|
|
|
Author:
|
|
| Name 1:
|
ستاره رضاخانی - Corresponding - Presenter
|
| Affiliation 1:
|
Shahid Beheshti university, Laser and Plasma Research Institute, Tehran, Iran
|
|
|
| Name 2:
|
|
| Affiliation 2:
|
|
|
|
| Name 3:
|
|
| Affiliation 3:
|
|
|
|
| Name 4:
|
|
| Affiliation 4:
|
|
|
|
| Name 5:
|
|
| Affiliation 5:
|
|
|
|
| Name 6:
|
|
| Affiliation 6:
|
|
|
|
| Name 7:
|
|
| Affiliation 7:
|
|
|
|
| Name 8:
|
|
| Affiliation 8:
|
|
|
|
| Name 9:
|
|
| Affiliation 9:
|
|
|
|
| Name 10:
|
|
| Affiliation 10:
|
|
|
|
Abstract
|
Original Research
|
|
introduction :
|
Autism spectrum disorder (ASD) affects how a person communicates, interacts, and experiences the world. Yet diagnosis still depends almost entirely on behavioral observation, which can be subjective, time-consuming, and delayed. A brain-based biomarker derived from a short resting-state fMRI scan could offer clinicians an objective second opinion, helping families get answers sooner. However, most connectivity studies treat the brain as a static network, computing a single correlation matrix that ignores two fundamental features of neural communication: direction and time. Brain regions influence each other asymmetrically, and these influences shift from moment to moment. Effective connectivity—the directed, time-varying influence between regions—may therefore capture ASD-related network dysfunction far better than conventional approaches.
|
|
Methods:
|
We propose a Spatial-Temporal Graph Transformer (STGT) that models the brain as a time-varying directed graph of 116 anatomically defined brain regions (AAL atlas). A temporal Transformer encoder learns long-range dependencies in regional BOLD time series, while Graph Attention Network (GATv2) layers infer time-resolved, directed edge weights between regions. The model is trained end-to-end on the multi-site ABIDE dataset (N = 1,112) in a multi-task framework that simultaneously predicts ASD diagnosis, ADOS total severity, and age. After training, attention weights are extracted to construct individual dynamic connectomes, giving the model direct neurobiological interpretability rather than operating as a black box.
|
|
Results:
|
The proposed STGT model, outperformed static functional connectivity baselines, achieving an AUC of 0.84 for ASD classification. Severity prediction correlated significantly with actual ADOS total scores (r = 0.45, p < .001). Age prediction yielded a mean absolute error of 1.8 years, demonstrating the model's sensitivity to developmental changes. Interpretability analyses revealed aberrant directed connections primarily involving the default mode and salience networks, with significant sex- and age-related effects after FDR correction (p < .05). The trained model and code were released as open-source resources.
|
|
Conclusion:
|
This work demonstrated that dynamic effective connectivity, learned through a graph transformer, can serve as a clinically interpretable and brain-based aid for ASD diagnosis and severity assessment. By capturing both the direction and dynamics of neural interactions, the framework moved beyond static connectivity toward a richer understanding of how brain networks reorganize in autism. Beyond ASD, the approach is generalizable to other brain mapping applications where understanding directed, time-varying network dynamics is essential.
|
| Attachment:
|
|
| Keywords:
|
FMRI, effective connectivity, graph transformer, autism, biomarker
|
| Key message:
|
|
| Do you want your abstract to be indexed, if this is selected as one of the top 100
articles?
|
