| Title:
|
Deep Learning–Driven Integration of Signed Graph Metrics and Temporal Dynamics for Biomarker Discovery in ADHD Using Resting-State fMRI
|
| Author(s):
|
1
|
| Subject:
|
Artificial Intelligence and Brain Mapping
|
|
|
|
Presenter Author:
|
|
| Name:
|
Rouzbh Fakhari
|
|
|
|
Author:
|
|
| Name 1:
|
Rouzbh Fakhari - Corresponding - Presenter
|
| Affiliation 1:
|
Ph.D. in Cognitive Science — Computational Neuroscience & AI, Institute for Cognitive Science & AI, 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 :
|
This work builds predictive models by leveraging signed topology and temporal dynamics. Prior signed-graph descriptive analyses (see Paper 1) indicate clinically relevant patterns in structural balance and negative-subnetwork TMH. Here, we develop an end-to-end deep learning pipeline that fuses signed graph features across time to derive interpretable, generalizable biomarkers for ADHD classification and symptom estimation.
|
|
Methods:
|
Data: ADHD-200 subset (n=620; 340 ADHD, 280 controls). Shared preprocessing and parcellation choices follow those in Paper 1 (see Supplement for shared parameters). Dynamic FC: sliding-window correlations (30–60 s) produced time-resolved signed adjacency matrices. For each window, signed-graph features (balance energy, triad proportions, negative-subnet TMH) were computed and combined with raw signed adjacency inputs. Model: a signed-graph neural network (signed-GNN) processes each window to produce per-window graph embeddings; embeddings are stacked and processed by a temporal transformer module (self-attention) to capture long-range dependencies; a classification head outputs ADHD vs control (end-to-end training). Baselines: SVM/RF on static FC and handcrafted metrics; ablation: remove signed edges branch / remove transformer. Training: nested cross-validation with site-aware splits to mitigate site effects. Interpretability: attention maps, GNN saliency, and SHAP analyses. Multimodal fusion (EEG–fMRI) evaluated in a subset (late fusion and joint embedding).
|
|
Results:
|
The GNN–transformer model significantly outperformed static baselines (p < 0.01): held-out accuracy ≈ 80% and improved AUC, vs ~70% for SVM on static FC and ~65% for static balance features. Ablation of signed edges or transformer produced marked performance drops. Incorporating EEG in the subset improved robustness and generalization. Interpretability analyses implicated frontal-parietal, default-mode, and cerebellar nodes and identified transient signed edges that strongly influenced predictions.
|
|
Conclusion:
|
Paper 2 demonstrates that fusing signed-graph topology with temporal deep learning yields predictive, interpretable biomarkers for ADHD. This manuscript complements Paper 1 by focusing on supervised prediction, ablation studies, multimodal fusion, and explainability. Future work will extend external validation, prospective longitudinal testing, and assessment of treatment-sensitive biomarker dynamics.
|
| Attachment:
|
|
| Keywords:
|
Attention Deficit Disorder with Hyperactivity; Magnetic Resonance Imaging; Brain Connectivity; Neural Networks, Computer; Biomarkers
|
| Key message:
|
|
| Do you want your abstract to be indexed, if this is selected as one of the top 100
articles?
|
