D-GATNet: Interpretable Temporal Graph Attention Learning for ADHD Identification Using Dynamic Functional Connectivity

Qurat Ul Ain 1,2 , Alptekin Temizel 2 , Soyiba Jawed 1 · 2026

Read D-GATNet: Interpretable Temporal Graph Attention Learning for ADHD Identification Using Dynamic Functional Connectivity by Qurat Ul Ain 1,2 , Alptekin Temizel 2 , Soyiba Jawed 1 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-26308.txt.

Introduction

Ain, Temizel, and Jawed propose D-GATNet, a graph model that flags ADHD from brain scans. The model uses dynamic functional connectivity, built by sliding a window over fMRI data. Brain regions form graph nodes, and link strengths form the edges. A graph attention network reads spatial links, while convolution and temporal attention read change over time. Attention weights show which regions and time spans drive each call, keeping the model open to inspection.

Catalog record

CREATORQurat Ul Ain 1,2 , Alptekin Temizel 2 , Soyiba Jawed 1
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE19350
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.26308

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

D-GATNet: Interpretable Temporal Graph Attention Learning for ADHD Identification Using Dynamic Functional Connectivity Thanks: Identify applicable funding agency here. If none, delete this. Qurat Ul Ain 1,2 , Alptekin Temizel 2 , Soyiba Jawed 1 Abstract Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder whose neuroimaging-based diagnosis remains challenging due to complex time-varying disruptions in brain connectivity. Functional MRI (fMRI) provides a powerful non-invasive modality for identifying functional alterations. Existing deep learning (DL) studies employ diverse neuroimaging features; however, static functional connectivity remains widely used, whereas dynamic connectivity modeling is comparatively underexplored. Moreover, many DL models lack interpretability. In this work, we propose D-GATNet, an interpretable temporal graph-based framework for automated ADHD classification using dynamic functional connectivity (dFC). Sliding-window Pearson correlation constructs sequences of functional brain graphs with regions of interest as nodes and connectivity strengths as edges. Spatial dependencies are learned via a multi-layer Graph Attention Network, while temporal dynamics are modeled using 1D convolution followed by temporal attention. Interpretability is achieved through graph attention weights revealing dominant ROI interactions, ROI importance scores identifying influential regions, and temporal attention emphasizing informative connectivity segments. Experiments on the Peking University site of the ADHD-200 dataset using stratified 10-fold cross-validation with a 5-seed ensemble achieved 85.18% ± 5.64 balanced accuracy and 0.881 AUC, outperforming state-of-the-art methods. Attention analysis reveals cerebellar and default mode network disruptions, indicating potential neuroimaging biomarkers. I Introduction Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder in children, characterized by inattention and hyperactivity [ 1 ] . Early diagnosis is crucial for reducing long-term adverse outcomes [ 2 ] . Recent advances in neuroimaging-based research, driven by machine learning (ML) and deep learning (DL), have improved automated ADHD classification [ 3 , 4 ] . Among neuroimaging modalities, functional magnetic resonance imaging (fMRI) has emerged as a promising auxiliary diagnostic tool. As a non-invasive technique with high spatial resolution and acceptable temporal resolution, fMRI is widely used to detect structural and functional brain alterations [ 5 ] [ 6 ] . Consequently, DL frameworks have been increasingly applied to fMRI data to enhance ADHD diagnosis. For example, Zhao et al. [ 7 ] proposed a dynamic graph convolutional network (dgCN) based on functional connectivity (FC) brain graphs, evaluated using 10-fold cross-validation (CV), and also interpreted potential ADHD biomarkers. Qiang et al. [ 8 ] introduced a Spatiotemporal Attention Autoencoder (STAAE) using resting-state temporal templates evaluated under 5 fold CV. Chen et al. [ 9 ] developed Att-AENet for FC based biomarker learning using leave-one-out-cross-validation (LOOCV) with interpretable region identification. Peng et al. [ 10 ] proposed Summation-based Synergic Artificial Neural Network (SSANN) with dual 3D-CNN branches for the fusion of sMRI and rs-fMRI data, evaluated using an 80:20 hold-out split. Wang et al. [ 11 ] proposed two approaches, ICA-CNN and correlation autoencoder, evaluated under hold out validation. Pei et al. [ 6 ] introduced Gaussian noise, Mix-up, and sliding-window augmentation and developed a CNN–GAT fusion model leveraging FC features, with the sliding window strategy demonstrating superior performance among the augmentation methods. Zhang et al. [ 12 ] proposed a Diffusion Kernel Attention Transformer framework, evaluated using an 80:20 data split, and performed interpretability analysis. Zhang et al. [ 13 ] proposed a model named BNC-DGHL, which incorporates CNN and hashing layers and evaluated under 5-fold CV. Hu et al. [ 14 ] introduced Source Free Semi-Supervised Transfer Learning (S3TL) with LSTM-based region of interest (ROI) features, evaluated using 10 fold CV. Gulhan and Özmen [ 15 ] applied 3D-CNNs to faLFF/ReHo measures evaluated with 5 fold CV. Despite this progress, robust ADHD diagnosis remains challenging. Many studies reporting strong performance rely on LOOCV [ 9 ] , which is computationally expensive and may produce high variance estimates that limit generalization. Moreover, although diverse features have been explored, static functional connectivity (FC) remains the most widely adopted representation [ 16 ] . However, static FC ignores the temporal dynamics of rs-fMRI, even though ADHD related abnormalities may emerge through time varying interactions. Although dynamic functional connectivity (dFC) has been examined in a few recent works [ 17 ] , its application to automated ADHD diagnosis remains comparatively underexplored, despite showing superiority over static measures in other psychiatric disorders [ 18 ] . In addition, most DL models lack interpretability, providing limited insight into the spatio-temporal patterns driving predictions. Therefore, we propose an interpretable resting-state functional MRI (rs-fMRI) driven ADHD classification framework leveraging dFC to capture both spatial and temporal abnormalities in functional brain networks. The main contributions are as follows: • We develop a temporal graph-based deep learning pipeline for ADHD identification using dFC representations extracted from rs-fMRI. • We model spatial interactions among brain regions through a Graph Attention Network (GAT) to learn discriminative ROI-to-ROI connectivity patterns. • We incorporate temporal convolution with an attention mechanism to capture time varying connectivity dynamics and highlight informative temporal windows. • We enhance clinical interpretability by exploiting graph and temporal attention weights to identify dominant ROI connections and dynamic brain states associated with ADHD. II Proposed Methodology In this study, we propose a D-GATNet (Dynamic Graph Attention Network) framework for automated ADHD identification using resting-state fMRI. Following preprocessing, dynamic functional connectivity (dFC) is extracted to model time varying brain connectivity patterns and capture complementary spatial and temporal information.The proposed architecture consists of five main modules: (i) Dynamic Connectivity Representation, (ii) Graph Construction, (iii) Spatial Graph Modeling, (iv) Temporal Dynamics Modeling, and (v) Classification. In addition, intrinsic interpretability is achieved through the learned graph and temporal attention weights, which highlight the most influential ROI connections and dynamic windows contributing to the final prediction. The overall framework is illustrated in. fig. 1 . Fig. 1: Overview of the proposed D-GATNet framework for ADHD classification. II-A Dynamic Connectivity Representation For each subject, the preprocessed rs-fMRI data are represented by an ROI-wise mean time series matrix \mathbf{X}\in\mathbb{R}^{L\times N} .where N=116 denotes the number of ROIs and L is the number of time points.To characterize time-varying functional interactions between brain regions, we compute dFC using Pearson correlation within a sliding window framework. Specifically, a temporal window of size w=40 with step size s=20 is applied to the ROI time series. For each window t , a connectivity matrix is obtained as defined in Eq. 1 . \mathbf{F}^{(t)}=\mathrm{corr}(\mathbf{X}_{t:t+w}) (1) where \mathbf{F}^{(t)}\in\mathbb{R}^{N\times N} denotes the functional connectivity graph at window t . Thus, each subject is represented as a sequence of dynamic FC matrices, as expressed in Eq. 2 . \{\mathbf{F}^{(t)}\}_{t=1}^{T} (2) where T is the total number of temporal windows. II-B Graph Construction Each FC matrix is interpreted as a brain graph, where ROIs correspond to nodes and connectivity strengths represent edges. Since correlation matrices are densely connected, thresholding is applied to retain the most informative connections and suppress noise. The adjacency matrix is defined according to Eq. 3 . \mathbf{A}^{(t)}_{ij}=\begin{cases}1,&|\mathbf{F}^{(t)}_{ij}|\geq\tau\\ 0,&\text{otherwise},\end{cases} (3) where \tau preserves the top 30% strongest functional links. Self loops are included to maintain node identity, producing a sparse graph representation suitable for graph neural learning. II-C Spatial Graph Modeling To learn discriminative spatial interaction patterns among ROIs, we employ a multi layer Graph Attention Network (GAT). Each ROI node is represented by a 64 dimensional embedding. Three GAT

Read the complete text