DW-DGAT: Dynamic Dual Graph Attention Network

• DW-DGAT is a graph-based model that fuses 1D, 2D, and 3D neuroimaging metrics to capture both region-level and subject-level dependencies.
• It employs a dual graph attention mechanism with dynamic weighting to effectively mitigate class imbalance and enhance early diagnosis of PD and AD.
• Experimental results on benchmark datasets demonstrate improved accuracy, sensitivity, and class separability compared to classical approaches.

The Dynamically Weighted Dual Graph Attention Network (DW-DGAT) is an advanced model for the integrated analysis of heterogeneous multi-metric neuroimaging and phenotypic data, designed to address challenges in early diagnosis of neurodegenerative diseases, particularly Parkinson’s disease (PD) and Alzheimer’s disease (AD). DW-DGAT combines a general-purpose fusion module for diverse structural data, a dual graph attention architecture capturing both region-level and subject-level dependencies, and a dynamically weighted loss mechanism for mitigating class imbalance. Its state-of-the-art performance is demonstrated on benchmark datasets for PD and AD, exhibiting superior accuracy, minority-class sensitivity, and class separability compared to classical approaches (Liang et al., 15 Jan 2026).

1. Problem Formulation and Motivation

Early diagnosis of neurodegenerative diseases is complicated by three fundamental issues: (a) neuroimaging data are high-dimensional and structurally heterogeneous, encompassing 1D regional statistics, 2D connectivity matrices, and 3D volumetric imaging; (b) prodromal stage differentiation (e.g., distinguishing healthy controls from early mild cognitive impairment or prodromal PD) requires both fine-grained regional features and global subject-level context; (c) available public datasets are highly imbalanced, with a minority of early-stage disease examples, exacerbating model bias and instability.

DW-DGAT addresses these by implementing:

• A general fusion mechanism that aligns and combines multi-metric data at the region-of-interest (ROI) level.
• A dual graph attention mechanism extracting micro-level (within-ROI) and macro-level (across-subject) representations.
• A dynamically weighted generator-classifier loss framework to stabilize training and enhance minority-class performance under class imbalance conditions (Liang et al., 15 Jan 2026).

2. General-Purpose Data Fusion for Heterogeneous Neuroimaging Inputs

DW-DGAT employs a three-stage procedure to transform and aggregate disparate metrics into a unified matrix for each subject:

• 1D metrics: Vectors such as ROI surface area and voxel count $$(\mathbf{v}_{\rm surf}^{(1D)}, \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R)$$ are fused elementwise:

$$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$

• 2D metrics: Connectivity matrices $(M^{(2D)} \in \mathbb{R}^{R \times R})$ are first min–max normalized, then $L_1$-row pooled:

$$\mathbf{u}^{(2D)} = \left[\|M_{1, \cdot}\|_1, \dots, \|M_{R,\cdot}\|_1 \right]^\top \in \mathbb{R}^R.$$

• 3D metrics: For each ROI $r$, DTI-derived volumes are reduced via barycenter-weighted coordinates $(\bar x_r, \bar y_r, \bar z_r)$, value at the barycenter $w_r$, mean $\bar w_r$, and maximum $\hat w_r$:

$$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$0

These are range-scaled and concatenated across ROIs.

The final fused ROI-by-feature matrix $$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$1 unifies all ROI-level metrics, facilitating subsequent graph-based modeling. This approach is adaptable to any combination of 1D, 2D, and 3D neuroimaging features (Liang et al., 15 Jan 2026).

3. Dual Graph Attention Architecture: Micro- and Macro-Level Modeling

Single Graph Attention (SGA; ROI Graph, Micro-Level)

• Nodes represent ROIs (90 in AAL-90).
• Edges are derived from pairwise Euclidean distances, forming a "centrality distance" measure.
• Pruning retains ROIs with strongest centrality, after which a Gaussian kernel similarity is added to features.
• The processed node features are embedded ($$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$2), prepended with a learnable CLS token and positional encodings, and passed through a 12-layer Vision Transformer (ViT). The final CLS token embedding $$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$3 forms the subject-level vector.

Global Graph Attention (GGA; Subject Graph, Macro-Level)

• Nodes correspond to all samples (subjects or timepoints), using ViT-derived features.
• Edge weights are defined via phenotype vector similarities:

$$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$4

• Two multi-head self-attention graph convolution (MHSA-GC) layers, each with $$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$5 heads, operate as follows:

$$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$6

with query, key, value projections per head, dynamic neighbor attention,

$$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$7

and aggregation across all heads and neighbors, followed by LayerNorm and GELU. A final fully connected layer reduces representation dimensionality.

This architecture ensures micro-level extraction of ROI connectivity/morphology patterns and macro-level modeling of inter-subject (phenotype, group) relations (Liang et al., 15 Jan 2026).

4. Dynamically Weighted Loss and Class Imbalance Handling

To robustly address class imbalance, DW-DGAT incorporates a Class Weight Generator (CWG) that mirrors DGAT’s subject-graph attention but is structured as $$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$8 separate streams (one per class mask). Each stream produces per-sample logits $$\mathbf{u}^{(1D)} = \mathbf{v}_{\rm surf}^{(1D)} \oslash \mathbf{v}_{\rm vox}^{(1D)} \in \mathbb{R}^R.$$9, stacked into $(M^{(2D)} \in \mathbb{R}^{R \times R})$0. Sample-class weights are computed via stabilized softmax: $(M^{(2D)} \in \mathbb{R}^{R \times R})$1 The DGAT classifier’s output logits $(M^{(2D)} \in \mathbb{R}^{R \times R})$2 are softmaxed to $(M^{(2D)} \in \mathbb{R}^{R \times R})$3; weighted cross-entropy loss is then calculated: $(M^{(2D)} \in \mathbb{R}^{R \times R})$4 The overall loss couples the weighted cross-entropy with standard SGA cross-entropy, and CWG training adds a negative entropy regularizer: $(M^{(2D)} \in \mathbb{R}^{R \times R})$5 This cooperative objective stabilizes optimization and sharpens minority class sensitivity. Stability-enhancing numerical shifts (subtracting max/min) and additive $(M^{(2D)} \in \mathbb{R}^{R \times R})$6 are critical for robust training (Liang et al., 15 Jan 2026).

5. Model Training Protocols and Implementation

DW-DGAT is implemented in PyTorch with CUDA acceleration on a single 24 GB GPU. Training follows:

• Adam optimizer, learning rate $(M^{(2D)} \in \mathbb{R}^{R \times R})$7,
• Dropout rate $(M^{(2D)} \in \mathbb{R}^{R \times R})$8 in ViT and MHSA-GC,
• ViT: 12 layers, $(M^{(2D)} \in \mathbb{R}^{R \times R})$9, 12 heads,
• GGA: $L_1$0 heads (PPMI), $L_1$1 (ADNI3),
• Batch size: 64,
• Epochs: 500,
• Seed: 231,
• Ten-fold cross-validation, with strict non-leakage across temporal/subject folds.

Key stabilizing tricks include pre-softmax normalization (subtracting extremal logits), numerical offsets ($L_1$2) on class weights, and post-attention LayerNorm. These are essential for consistent convergence, especially under highly imbalanced conditions.

The end-to-end workflow proceeds from multimodal MRI/DTI preprocessing (PANDA + FSL) through fusion, SGA (ViT), GGA (MHSA-GC), dynamic weighting, and inference (Liang et al., 15 Jan 2026).

6. Experimental Evaluation and Comparative Analysis

DW-DGAT was evaluated on:

• PPMI (PD): 316 subjects (69 HC, 72 PRO, 175 PD) yielding 636 samples across three timepoints.
• ADNI3 (AD): 310 subjects (163 CN, 118 EMCI, 29 AD), 464 samples.

Performance metrics included accuracy (ACC), balanced accuracy (BA), F1, specificity (SPE), and AUC.

Main results (mean ± std, 10-fold cross-validation):

Dataset	ACC (%)	BA (%)	F1 (%)	SPE (%)	2nd-Best Baseline (ACC)	ACC Gain (%)
PPMI	74.56 ± 5.99	59.31 ± 8.73	70.57 ± 7.31	79.66 ± 4.36	66.99 (ViT-small)	+7.57
ADNI3	68.65 ± 4.35	66.18 ± 9.48	66.79 ± 5.23	83.09 ± 4.74	64.03 (ViT-small)	+4.62

Ablation results indicate that each module adds performance: MLP baseline (~63% ACC, PPMI), +Data Fusion (+2.36%), +SGA (+4.40%), +GGA (+8.65%), complete CWG reaches 74.56%. t-SNE analysis shows tighter clustering of classes and improved minority-class separation. ROC curves confirm superior trade-off over all baselines (Liang et al., 15 Jan 2026).

7. Analysis, Limitations, and Future Directions

DW-DGAT’s effectiveness is attributed to:

1. The fusion module’s ability to harness complementary statistics from 1D, 2D, and 3D modalities,
2. Dual-attention across ROIs and subjects, yielding both local and global contextual embeddings,
3. Adaptive class weighting that explicitly addresses dataset imbalance and sharpens convergence stability.

Noted limitations include potential bias from unnormalized 3D ROI volumes (risk of overweighting larger regions), and substantial computational demands (notably temporal $L_1$3 complexity for GGA layers). Prospective enhancements involve ROI-specific volume adjustment, adoption of lightweight transformer architectures, and transfer to broader multi-modal diagnostic domains (Liang et al., 15 Jan 2026).

• DW-DGAT: Dynamically Weighted Dual Graph Attention Network for Neurodegenerative Disease Diagnosis (Liang et al., 15 Jan 2026)
• DAGCN: Dual Attention Graph Convolutional Networks (Chen et al., 2019)