Enhancing depression diagnosis with augmented brain signal driven decorrelated graph neural networks

Abstract

Background

Major Depressive Disorder (MDD) is a leading global neuropsychiatric disorder, requiring precise diagnosis for effective intervention. Developing accurate diagnostic models for MDD remains a critical but challenging task. This study introduces a graph-based deep learning framework that addresses the issue of limited training data and facilitates robust training for identifying MDD across diverse episode patterns.

Methods

We introduce Brain Augmented-Decorrelated Network (BrainADNet), a framework designed to address data scarcity by augmenting brain signal inputs. BrainADNet builds upon the Skip-Graph Convolutional Network to aggregate informative multi-layer features, enriching its representational capacity. Recognizing the clinical relevance of demographic factors such as age, education, and gender in depression, we incorporate these attributes into the training process and examine their effect on diagnosis. To further improve feature diversity and reduce overfitting, we use a decorrelation regularizer to the model training. This encourages GCN embeddings to learn complementary, non-redundant representations from input graphs.

Results

As far as we are aware, the framework surpasses existing models in accurately identifying MDD cases across depressive stages. We present a detailed ablation study demonstrating the contribution of each component to diagnostic precision. Our study highlights the top-10 brain regions influential in diagnosing MDD in males and females, addressing a crucial gap in understanding gender-specific neural mechanisms. We also uncover distinct patterns in latent-space brain connectivity, derived from GCN embeddings, between individuals experiencing single versus multiple depression episodes.

Conclusions

This study underscores the potential of graph methods to advance diagnostic precision for MDD. By integrating gender-specific and stage-wise insights, our framework equips medical professionals and researchers to design personalized and targeted therapeutic strategies, offering transformative implications for patient care.

Plain language summary

Major Depressive Disorder (MDD) is a serious mental health condition characterized by a persistently low mood lasting at least two weeks, often accompanied by a range of symptoms that differ from person to person. MDD significantly affects daily life, relationships, brain health, and in severe cases, it may lead to suicidal attempts. Addressing the stigma around MDD and improving diagnosis are essential for timely and effective treatment. In this study, we introduce a graph-based method to accurately detect MDD, identify its stages, and extract distinct brain regions involved in both men and women. In addition, our approach sheds light on how brain functional connectivity changes with MDD progression, helping to refine diagnoses and support personalized treatment strategies.

Barman et al. present a graph machine-learning framework for identifying individuals with depression and their severity stages. The model improves classification accuracy, mitigates feature redundancy, and reveals gender-specific brain regions and stage-wise functional connectivity changes.

Introduction

The World Health Organization’s (WHO) Mental Health Action Plan 2013-2030 and Mental Health Gap Action Programme (mhGAP) prioritize addressing mental disorders, particularly depression and self-harm, including suicide prevention¹. Major depressive disorder (MDD), a complex condition influenced by biological, environmental, and psycho-social factors, was the third largest contributor to the global disease burden in 2008 and is projected to become the leading cause by 2030^2,3. According to the Diagnostic and Statistical Manual of Mental Disorders (DSM-V), diagnosis requires persistent symptoms like low mood, sleep disturbances, difficulty concentrating, and thoughts of self-harm, all of which severely impact social and personal life⁴. Currently, MDD affects 3.8% of the global population, with higher prevalence in women (6%) compared to men (4%), and in older adults (5.7% of those over 60)⁵. Women are ~50% more likely to experience MDD than men. Recurrent MDD (rMDD), a more severe and persistent form, involves repeated depressive episodes and has been recognized in the WHO’s International Classification of Diseases (ICD-11)⁶. Alarmingly, individuals with a single depressive episode face a 50% risk of recurrence, which increases to 80% after two episodes⁷. Despite the availability of effective treatments, over 75% of people in low- and middle-income countries remain untreated due to limited resources and the stigma surrounding mental illness. Misdiagnosis or delayed diagnosis of MDD is common, as its symptoms are often mistaken for normal emotional fluctuations. Accurate and timely identification of MDD is crucial for delivering effective treatments, reducing the risk of recurrence, and alleviating the long-term psychological, social, and economic burden on healthcare systems.

Among neuroimaging techniques, functional magnetic resonance imaging (fMRI) has emerged as a critical tool for capturing neuronal activity in specific brain regions^8,9. Studies reveal that depression significantly impacts structural neuroplasticity, altering both structural and functional connectivity in key brain areas such as the prefrontal cortex, hippocampus, amygdala, and gray and white matter^9–11. Early investigations into classifying MDD patients versus Healthy Controls (HCs) have uncovered distinct alterations in functional brain networks, especially in resting-state fMRI (rs-fMRI) data^12–14. The default mode network (DMN), typically active during rest, exhibits functional connectivity and heightened metabolic activity in healthy individuals^15,16. However, large-scale analyses reveal marked reductions in DMN connectivity in recurrent MDD (rMDD) patients (t = −3.737, p = 0.0002, d = −0.326), while only minimal alterations are observed in first-episode MDD patients (t = −0.914, p = 0.361, d = −0.076)¹⁷. These findings underscore the profound effects of depression on brain connectivity and suggest that DMN disruptions may serve as a biomarker for recurrent MDD, offering valuable insights for diagnosis and intervention.

Traditional machine learning and deep learning based fMRI classification methods, such as support vector machines (SVM), logistic regression, and decision trees, rely on feature extraction but often fail to capture the critical topological relationships between brain regions^18,19. Although deep learning techniques have shown promise in medical imaging, they are not designed to handle the intricate connectivity architecture of the human brain^20–22. This limitation leads to suboptimal performance in disease classification, as accurately mapping the complex connectivity of the brain network is essential for understanding neurological disorders. Graph-based framework offers a solution by modeling brain regions as nodes and their functional connections as the edges^23–25. Graph neural networks (GNNs) build on this foundation by integrating graph structures with neural networks, enabling the simultaneous processing of node attributes and their interconnections. This approach allows GNNs to effectively capture complex brain network patterns, significantly enhancing brain disorder classification and cognitive state prediction^26,27. Recent works employ GCNs to capture topological properties of brain connectivity, achieving favorable performance in identifying MDD^28,29. Additionally, GNNs have been investigated for interpretable diagnostics via linear models leveraging global network measures, as well as for multimodal integration and multi-task learning, showing potential to enhance diagnostic precision and adaptability^30–32.

Existing graph-based approaches face a critical limitation in generalizing under limited annotated data. Generative architectures, such as variational autoencoders and diffusion-based models, have shown strong potential for data augmentation in various domains. However, these methods face several critical limitations when applied to brain connectivity analysis. These models often rely on large and diverse sets of unlabeled data, require substantial computational resources, and typically struggle to preserve biologically meaningful connectivity patterns^33,34. Again, diffusion-based methods often lack domain-specific inductive biases and involve complex optimization procedures, making their direct application to brain graph analysis challenging³⁵. In parallel, semi-supervised and self-supervised learning are two active research paradigms for learning under low-label conditions. In semi-supervised learning, the model is trained to minimize error on the labeled data while enforcing consistency on the unlabeled data. In contrast, self-supervised learning focuses on learning invariant features by solving pretext tasks, followed by fine-tuning on labeled data. Both paradigms typically require enough task-specific data to achieve desirable performance and are often computationally intensive. We argue that effective augmentation strategies for labeled data can substantially improve the performance of any learning paradigm.

Motivated by this, we propose BrainADNet, an efficient and interpretable framework to diagnose individuals with depressive disorder and identify their stages. To address the issue of data scarcity, we first implement a sliding-window technique on the original time-series brain signals, also known as blood oxygenation level dependent (BOLD) signals, as a data augmentation step. This approach not only mitigates the issue of limited data but also standardizes the BOLD time series to a fixed dimension across all subjects from different sites. In this work, unlike conventional approaches that focus solely on brain signals, BrainADNet integrates demographic information, including age, gender, and years of education, along with brain connectivity features, providing a more comprehensive understanding of how this personal information influences the training and performance of the model. Following ref. ³⁶, a core component of BrainADNet is Skip-GCN, which is designed to mitigate overfitting while enhancing classification accuracy and reliability. The dynamic weight fusion block in Skip-GCN aggregates information from each layer. To mitigate overfitting, which is a common issue when training neural networks on limited data, we use a decorrelation regularizer along with the traditional GCN loss function. By reducing correlations between hidden units, the regularizer promotes independent and diverse graph representations, improving both training efficiency and model performance. Since graph classification in neuroimaging is typically formulated as a supervised learning problem³⁷, we present our results in a supervised setting. However, the techniques we propose are not limited to supervised frameworks. In preliminary evaluations, we find that the same augmentation strategies yield consistent benefits when incorporated into SSL frameworks, highlighting the flexibility and broad applicability of our approach.

We validate the effectiveness of BrainADNet using rs-fMRI data from the Rest-meta-MDD consortium (https://rfmri.org/REST-meta-MDD)³⁸, which contains data from 25 research groups across 17 hospitals in China¹⁷. Extensive experiments on the benchmark dataset demonstrate that BrainADNet consistently outperforms existing methods in classifying depressive disorders. We also investigate gender-specific biomarkers and the stage-wise evolution of functional connectivity networks associated with depression.

Materials and methods

Data description

The dataset used in this study comprises preprocessed rs-fMRI from 1300 patients diagnosed with MDD and 1128 HCs¹⁷. All participants provided written informed consent. The study from which the dataset was derived was approved by the Institutional Review Board of the Institute of Psychology, Chinese Academy of Sciences (No. H21102)³⁹. All the primary information, such as diagnosis, age, sex, and education, as well as the Hamilton Depression Rating Scale (HAMD) corresponding to each participant, is also provided in the dataset. After a comprehensive quality assessment, the dataset was refined to include 848 individuals diagnosed with MDD and 794 HCs from 17 distinct hospitals/sites. In the multiclass classification framework, we aimed to differentiate between three distinct groups: First-episode drug-naive (FEDN) patients, representing individuals experiencing their first depressive episode without prior psychiatric medication; rMDD patients, characterized by individuals with multiple depressive episodes; and HCs with no history of psychiatric disorders. Participants for the multiclass classification analysis were drawn from nine hospitals/sites (Sites 7–9, 11, 14, and 19–21, 23), comprising a total of 937 individuals: 529 HCs, 219 FEDN patients, and 189 rMDD patients. Detailed demographic and clinical information of the participants are provided in the sheet entitled Dataset_details in the Supplementary Data. Additionally, our analysis categorized participants into three pairwise comparison groups to further investigate distinctions: MDD vs. HC, rMDD vs. HC, and FEDN vs. HC. In the MDD category, we include both rMDD and FEDN individuals. The data statistics for binary classification are as follows: for the MDD vs. HC classification, the dataset comprises 1570 participants from 16 sites (Sites 1, 2, 7–11, 13–15, 17, 19–23), including 756 HCs and 814 individuals diagnosed with MDD. For the FEDN vs. HC classification, data are obtained from 5 sites (Sites 8, 9, 14, 20, 23), comprising 598 participants (379 HCs and 219 FEDN). For the rMDD vs. HC classification, the dataset consists of 615 participants from 6 sites (Sites 7, 9, 11, 19–21), including 426 HCs and 189 rMDD patients.

Data preprocessing

To eliminate variability, the T1-weighted rs-fMRI images were preprocessed using a standardized preprocessing pipeline, data processing assistant for resting-state fMRI (DPARSF) at each local participating site⁴⁰. The following preprocessing steps were performed: (i) Slice timing correction: to synchronize all slices to a reference time point, (ii) Realignment: to correct the functional data for a movement that has occurred during scanning, (iii) removing covariates: used data with global signal regression, (iv) Spatial normalization: to ensure consistent anatomical coordinates across all the subjects, (v) Co registration: used to put the structural and functional data in alignment, (vi) Band-pass filtering: performed using a 0.01-0.1 Hz filter, (vii) Normalization: normalized to a symmetric template and finally smoothened to enhance the signal and cancel out any noise. Following preprocessing, time series BOLD signals corresponding to specific brain atlases are extracted. Subsequently, poor-quality and incomplete BOLD signals are also excluded across the various atlases: anatomical automatic labeling (AAL) - 116 ROIs⁴¹, Dosenbach - 160 functional ROIs⁴², Craddock - 200 ROIs⁴³, where ROI refers to the region of interest.

Ethics statement

In this work, we propose a graph-based machine learning framework for the diagnosis of depression, with the goal of improving mental health outcomes. We utilize a controlled access dataset³⁸ (http://rfmri.org/REST-meta-MDD) and implement model development and evaluation procedures to ensure scientific integrity. This research was conducted in accordance with established ethical guidelines and reflects our commitment to inclusion, equity, and responsible research practices.

Graph representation learning

Here, the graph is denoted as $\mathcal{G}=\left(\mathcal{V},\mathcal{E},\overline{W}\right)$, where $\mathcal{V}$ is the set of nodes represented by brain regions or ROIs, $\mathcal{E}$ is the set of edges which signify the connectivity pattern among the brain regions, and $\overline{W}$ is the weighted adjacency matrix which identifies how strongly they are connected. The original BOLD signals are denoted as $\overline{B}\in {\mathbb{R}}^{N\times T}$, where N is the number of brain regions and T is the number of time steps involved in the scanning process. The formula to calculate Pearson’s correlation coefficient can be expressed as:

$$R_{ij}=\frac{{\sum }_{k=1}^T\left(p_{i,k}-{\overline{p}}_i\right)\left(p_{j,k}-{\overline{p}}_j\right)}{\sqrt{{\sum }_{k=1}^T{\left(p_{i,k}-{\overline{p}}_i\right)}^2}\sqrt{{\sum }_{k=1}^T{\left(p_{j,k}-{\overline{p}}_j\right)}^2}}\in {\mathbb{R}}^{N\times N}$$ 1

Here, p_i,k and p_j,k represent the BOLD values for ROIs i and j at time point k. Their mean values, ${\overline{p}}_i$ and ${\overline{p}}_j$, are calculated as ${\overline{p}}_i=\frac{1}{T}{\sum }_{k=1}^T{p}_{i,k}$ and ${\overline{p}}_j=\frac{1}{T}{\sum }_{k=1}^T{p}_{j,k}$. To compute Fisher’s Z-transform on each correlation matrix R, we define: $Z_{ij}^{\left(f\right)}=\frac{1}{2}ln\left(\frac{1+R_{ij}}{1-R_{ij}}\right)=\mathrm{artanh}\left(R_{ij}\right)$. Fisher’s Z-transformation stabilizes the variance of Pearson correlation coefficients and transforms them into an approximately normal distribution, thereby enabling the application of parametric statistical methods with greater reliability. The final KNN graph $\mathcal{G}=\left(\mathcal{V},\mathcal{E},\overline{W}\right)$ can be defined as: $\mathcal{E}=\left\{\left\{v_i,v_j\right\}\mid v_j\in \mathrm{KNN}\left(v_i,\kappa \right),v_i\ne v_j\right\}$, where KNN(v_i, κ) represents the set of the κ nearest neighbors of v_i. The weighted adjacency matrix is defined by:

$${\overline{W}}_{ij}^{\left(\mathrm{knn}\right)}=\left\{\begin{array}{cc}\mathrm{Sim}\left(v_i,v_j\right)\hfill & \mathrm{if}\left\{v_i,v_j\right\}\in \mathcal{E},\hfill \\ 0\hfill & \mathrm{otherwise}.\hfill \end{array}\right)$$ 2

${\overline{W}}_{ij}$ captures the similarity of each node with its neighboring nodes. This enforces a sparse graph structure by preserving only the most relevant local connections for each node, thereby enhancing structural fidelity and reducing spurious links in the functional brain network.

Graph convolutional network

GNNs are a powerful tool for learning representations of graph-structured data. A specific type of GNN, the graph convolutional network (GCN), operates under the principle of message passing, where node features are updated by aggregating features from neighboring nodes.

For a given node $v\in \mathcal{V}$, GCNs learn the node embedding $Z_v^{\left(\ell \right)}\in {\mathbb{R}}^{d_n}$ at layer ℓ by aggregating messages from its neighbors $\mathcal{N}\left(v\right)$. Specifically, the embedding is updated by first aggregating the embeddings of neighboring nodes from the previous layer and then combining this aggregated message with the node’s own embedding from the previous layer. This process is formally expressed as:

$$Z_v^{\left(\ell \right)}=\mathrm{COM}\left(Z_v^{\left(\ell -1\right)},\mathrm{AGG}\left(\left\{Z_u^{\left(\ell -1\right)}:u\in \mathcal{N}\left(v\right)\right\}\right)\right).$$ 3

Here, AGG denotes an aggregation function (e.g., mean or sum), and COM denotes a combination function (e.g., concatenation followed by a linear transformation and nonlinearity). If self-loops are included in the graph, the node v may be part of its own neighborhood, i.e., $v\in \mathcal{N}\left(v\right)$.

GCNs introduce a layer-wise propagation rule that refines node features through a series of transformations and nonlinear activations. Let $A\in {\mathbb{R}}^{N\times N}$ be the adjacency matrix and $X\in {\mathbb{R}}^{N\times n}$ the input feature matrix. Starting with Z⁽⁰⁾ = X, the update rule for each GCN layer is:

$$Z^{\left(\ell \right)}=\sigma \left({\tilde{D}}^{-\frac{1}{2}}\tilde{A}{\tilde{D}}^{-\frac{1}{2}}{Z}^{\left(\ell -1\right)}{W}^{\left(\ell \right)}\right),$$ 4

where $\tilde{A}=A+I$ is the adjacency matrix with added self-loops, $\tilde{D}$ is the corresponding degree matrix with ${\tilde{D}}_{ii}={\sum }_j{\tilde{A}}_{ij}$, $W^{\left(\ell \right)}\in {\mathbb{R}}^{d_{\ell -1}\times d_{\ell }}$ is the trainable weight matrix, and σ( ⋅ ) is a nonlinear activation function. The output of the final GCN layer is $Z^{\left(L\right)}\in {\mathbb{R}}^{N\times d}$, where d is the embedding dimension. In graph classification tasks, node-level embeddings are aggregated into a graph-level representation using a readout (pooling) layer. For a given graph G, average pooling is expressed as:

$$Z_G=\frac{1}{\mid V\mid }\sum _{v\in V}{Z}_v^{\left(L\right)},$$ 5

where $Z_G\in {\mathbb{R}}^d$ is the graph-level representation. Other pooling methods can also be applied in the readout layer, such as max pooling, defined as $Z_G={\max }_{v\in V}{Z}_v^{\left(L\right)}$, where the $\max$ operator is applied element-wise across all node features, and sum pooling, defined as $Z_G={\sum }_{v\in V}{Z}_v^{\left(L\right)}$. These pooling strategies enable the model to produce a compact vector representation of the entire graph. This vector is then passed through a fully connected layer for classification.

In this study, we adopt ChebyNet for extracting higher-order graph representations⁴⁴. ChebyNet employs spectral graph convolution based on the graph Fourier transform, enabling localized filtering of node features using polynomial approximations. This approach extends the receptive field beyond 1-hop neighborhoods. Let the normalized graph Laplacian be decomposed as L = UΛU^⊤, where $U\in {\mathbb{R}}^{N\times N}$ is the matrix of eigenvectors and Λ = diag(λ₀, …, λ_N−1) is the diagonal matrix of eigenvalues. A spectral filter defined as f_θ(Λ) = diag(θ) enables convolution with a signal $p\in {\mathbb{R}}^N$ using the operation f_θ*p = Uf_θ(Λ)U^⊤p. To avoid the computational cost of eigendecomposition, the filter is approximated using a K-order Chebyshev polynomial expansion⁴⁴: $f_{{\theta }^{\prime }}*p\approx {\sum }_{k=0}^K{\theta }_k^{\prime }{T}_k\left(\tilde{L}\right)p,$ where ${\theta }^{\prime }\in {\mathbb{R}}^K$ is a vector of trainable Chebyshev coefficients, $T_k\left(\tilde{L}\right)$ is the Chebyshev polynomial of order k, and $\tilde{L}=\frac{2}{{\lambda }_{\max }}L-I$ is the scaled Laplacian. The Chebyshev polynomials are recursively defined as: $T_0\left(\tilde{L}\right)=I,T_1\left(\tilde{L}\right)=\tilde{L},T_k\left(\tilde{L}\right)=2\tilde{L}{T}_{k-1}\left(\tilde{L}\right)-T_{k-2}\left(\tilde{L}\right).$ In ChebyNet, starting from the input node features Z⁽⁰⁾ = X, the embedding at layer ℓ is updated as

$$Z^{\left(\ell \right)}={\sum }_{k=0}^K{\theta }_k^{\left(\ell \right)}{T}_k\left(\tilde{L}\right)Z^{\left(\ell -1\right)},$$ 6

where ${\theta }_k^{\left(\ell \right)}$ are the trainable parameters for each polynomial order. This formulation allows each node to aggregate information from nodes within its K-hop neighborhood, enabling more expressive representations. In this study, we use global mean pooling in the readout layer, followed by a fully connected linear layer for classification.

BrainADNet framework

Figure 1 illustrates the proposed framework of our study. Our first step towards feature extraction is data augmentation, where each original BOLD signal is segmented into multiple fixed-length time series. This is achieved through a sliding-window technique, with a window length of W = 60 and a stride of S = 30 for ensuring consistent time series BOLD signals across all the participants. This data augmentation strategy addresses the limited availability of brain signals by generating multiple signals from a single input. This augmented time series data is then used to derive the brain graph, where initially we calculate the Pearson’s correlation coefficients between every pair of ROIs across all the subjects using Equation (1). After applying Fisher’s z-transformation to stabilize the correlation coefficients, we incorporate demographic attributes to derive the final graph features. The KNN graph is constructed using Equation (2) on Fisher’s z-transformed features, where we set κ = 10. The application of KNN is pivotal for preserving the local connectivity structure of the data, as it captures the most relevant relationships between nodes while mitigating the impact of noise and sparsity. All the augmented graphs and features are labeled according to their original BOLD signals. Finally, we train the Skip-GCN model using a decorrelation loss function along with the traditional Cross-Entropy (CE) loss, effectively preventing overfitting and enhancing classification accuracy.

Fig. 1. Overall pipeline of BrainADNet for classifying depressed individuals across different stages and healthy controls.

a, b The original BOLD signal is segmented using a sliding window (W = 60, S = 30) to generate multiple augmented time-series segments. c Pearson’s correlation matrices are computed from each segment, followed by Fisher’s Z-transformation, and demographic information is early-fused with the resulting connectivity matrices to define graph features. d Graphs are constructed using a KNN algorithm (κ = 10). The learned features and graphs are then processed by Skip-GCN to classify first-episode drug-naive MDD patients, recurrent MDD patients, and healthy controls within a unified framework.

Skip-GCN

The framework of the GCN model with skip connections between layers is shown in Fig. 2. Specifically, we adopt a ChebyNet architecture composed of three Chebyshev (Cheb) convolutional layers. Each Cheb layer sequentially processes node features from the previous layer to compute updated representations. Within each layer, the input features are passed through a series of operations, including a ChebConv layer, batch normalization, ReLU activation, and dropout.

Fig. 2. Skip-GCN architecture used in our experiment.

The model consists of three convolutional layers with ReLU as the activation function, followed by a dropout layer. The output of each layer is connected to the subsequent layers via a skip connection, ensuring information flow across the network. The dynamic weight fusion block applies attention to each layer output and combines them to generate the final embeddings, with attention parameters as learnable weights in the architecture. The readout layer aggregates the fused embeddings to generate a graph-level representation. The fully connected linear layer utilizes this representation to determine the prediction score for classification.

Skip connections offer several advantages: It mitigates the vanishing gradient problem, enabling deeper architectures to train effectively; improve gradient flow across layers; and enhance feature reuse by directly combining low-level and high-level representations. To promote efficient feature aggregation and improve convergence, the output of each layer is added to the output of the next layer through a skip connection. Then, the Dynamic Weight Fusion Block aggregates outputs from all layers using trainable attention parameters within the neural network. This block applies learnable scalar weights to the outputs of each convolutional layer and computes a weighted sum to obtain the fused node embedding Z_fused as:

$$Z_{\mathrm{fused}}=\sum _{\ell =1}^L{a}_{\ell }{Z}^{\left(\ell \right)},$$ 7

where L is the total number of layers, a_ℓ are the learnable weights assigned to each Cheb layer output, and Z^(ℓ) denotes the output from the ℓ^th Cheb layer. The graph-level representation Z_G is obtained using a global mean pooling operation through the READOUT layer. Finally, the predicted target label $\tilde{y}$ is calculated using a fully connected linear layer, followed by an activation function σ (e.g., softmax or sigmoid):

$$\tilde{y}=\sigma \left({\theta }_{\mathrm{fc}}{Z}_{\mathrm{G}}+b_{\mathrm{fc}}\right),$$ 8

where ${\theta }_{\mathrm{fc}}\in {\mathbb{R}}^{C\times d}$ is the weight matrix of the fully connected linear layer, $b_{\mathrm{fc}}\in {\mathbb{R}}^C$ is the bias vector, and C is the number of output classes.

Decorrelation for generalization

To improve model generalization, it is important to learn representations that are both diverse and non-redundant. Minimizing activation correlation is a well-established strategy to achieve this^45,46. Following⁴⁷, we incorporate decorrelation regularization into GCN-based graph classification to enhance representation learning for fMRI-based depression diagnosis.

Consider a hidden layer in the GCN model, where the activations for sample i in a batch of size B are represented by ${\mathbf{z}}^{\left(i\right)}\in {\mathbb{R}}^d$. To quantify the correlation between neurons n₁ and n₂, we compute the covariance matrix defined as

$$C_{n_1,n_2}=\frac{1}{B}\sum _{i=1}^B\left(z_{n_1}^i-{\mu }_{n_1}\right)\left(z_{n_2}^i-{\mu }_{n_2}\right)$$ 9

where ${\mu }_{n_1}=\frac{1}{B}{\sum }_{i=1}^B{z}_{n_1}^i$, is the mean of activation n₁ over the batch. No additional normalization is applied to the covariance matrix, as it is inherently computed from mean-centered features. To suppress inter-feature redundancy, we minimize the squared Frobenius norm of the off-diagonal elements of the covariance matrix

$${\mathcal{L}}_{\mathrm{Decorr}}={∥C∥}_F^2-{∥\mathrm{diag}\left(C\right)∥}_2^2$$ 10

where ∥ ⋅ ∥_F is the Frobenius norm, and diag( ⋅ ) extracts the diagonal of C. By subtracting the diagonal norm, the loss specifically targets inter-neuron correlations without penalizing individual activation variances. This encourages feature decorrelation while preserving individual variances, helping the network discover a richer and more discriminative feature space. The computational cost of computing the batch covariance matrix is $\mathcal{O}\left(B{d}^2\right)$. In our case, B = 32 and d = 128. This is computationally marginal and does not dominate the overall training complexity of the GCN model.

BrainADNet objective

In our framework, the decorrelation loss is applied to the output of the READOUT layer, which aggregates node-level embeddings from the final Skip-GCN layer into graph-level embeddings $Z_{\mathrm{G}}\in {\mathbb{R}}^{B\times d}$. Minimizing the covariance promotes diversity in the learned embeddings by reducing redundancy across embedding dimensions. Through backpropagation, this regularizer influences earlier layers as well, guiding the network toward learning more discriminative and low-redundancy representations.

To this end, we optimize the following composite loss function:

$${\mathcal{L}}_{\mathrm{total}}=\frac{1}{B}\sum _{i=1}^B\left[-\left(y_i\log \left({\tilde{y}}_i\right)+\left(1-y_i\right)\log \left(1-{\tilde{y}}_i\right)\right)\right)\left(+\lambda \left({∥C\left(Z_{\mathrm{G}}\right)∥}_F^2-{∥\mathrm{diag}\left(C\left(Z_{\mathrm{G}}\right)\right)∥}_2^2\right)\right]+\gamma {∥\mathrm{\Theta }∥}_2^2$$ 11

Here, ${\tilde{y}}_i$ is the predicted probability for the i-th sample, y_i is the ground truth label, Θ represents model parameters, λ controls the contribution of the decorrelation regularizer, and γ is the weight decay coefficient.

Results

In this section, we present a comprehensive analysis of the results, highlighting both the classification performance and the underlying neurobiological insights. Beyond classifying individuals across different stages of depression, we also investigate the influential brain regions that contribute to the diagnosis of depression in both male and female individuals. Furthermore, we examine the evolution of brain functional connectivity with varying levels of depression severity, providing a deeper understanding of how these changes correlate with the progression of the disorder. We also conduct a series of comprehensive ablation studies to evaluate the contribution of each component of our method to the overall classification performance. For the ablation study and subsequent analysis, we consider the AAL atlas.

Evaluation metric

The performance metrics considered for classification include accuracy, F1 score, and area under curve (AUC). The calculation methods are as follows:

$$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}\mathrm{F1}\mathrm{Score}=\frac{2*\mathrm{Precision}*\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$

Here, $\mathrm{Recall}=\frac{TP}{TP+FN}$, $Precision=\frac{TP}{TP+FP}$, TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative. AUC is a performance metric that evaluates the ability of classification model to distinguish between the classes. It represents the degree of separability achieved by the model. The higher all the metrics, the better the model is at correctly classifying instances of different classes. We present all three metrics throughout our study to demonstrate the results.

Baselines and training details

We benchmark the proposed BrainADNet framework against a comprehensive set of state-of-the-art baseline models spanning four categories: classical machine learning, deep learning, generative augmentation, and self-supervised learning (SSL). The evaluation of classical machine learning, deep learning and generative models is conducted using tenfold cross-validation, with performance measured in terms of Accuracy, AUC, and F1 Score. To assess cross-institution generalization, we also perform leave-one-site-out (LOSO) evaluation. For comparisons against SSL baselines, we follow the training strategy presented in ref. ⁴⁸. Hyperparameters for each model are tuned based on the implementation details provided in their respective original papers. To ensure consistency, we use a fixed learning rate of 0.001, a weight decay of 5 × 10⁻⁴, and train each model for 50 epochs. The details of the baseline models and implementation settings are as follows:

• Classical machine learning baselines: We include SVM, Random Forest (RF), and Logistic Regression (LR), trained on the widely used “vectorized correlation matrix” feature^49,50. Specifically, for each subject, Pearson correlation coefficients are computed between resting-state fMRI time series from all ROI pairs. The upper-triangular (non-redundant) values are flattened into one-dimensional vectors and stacked to form a feature matrix $X_f\in {\mathbb{R}}^{P\times f}$, where f denotes the number of unique ROI-to-ROI connections and P is the number of subjects. Hyperparameters (e.g., kernel type and regularization strength) are selected via grid search within the training folds.

• Deep learning baselines: We benchmark against several graph- and transformer-based models, including GCN⁵¹, GAT⁵², GIN⁵³, TGCN⁵⁴, BrainGNN⁵⁵, BrainNetCNN⁵⁶, PopGCN⁵⁷, and BNT⁵⁸. All models are implemented from their official repositories and configured according to their original specifications. For GCN, GAT, and GIN, we adopt a three-layer architecture with hidden dimensions of 128, 64, and 32, respectively, and ReLU activations.

• Generative augmentation baselines: We evaluate data augmentation strategies using generative models, namely VAE⁵⁹ and VAE-GAN⁶⁰. Each model is trained following its original implementation to generate 100% new synthetic fMRI samples, effectively doubling the original training set. A GCN classifier is then trained on the augmented dataset and evaluated for classification performance.

• Self-supervised GNN baselines: We include two state-of-the-art SSL methods: GATE⁶¹, a temporal contrastive learning framework, and GCDA⁴⁸, a diffusion-based contrastive learning framework. For both, pretraining is conducted on the largest acquisition site (Site 20), followed by fivefold cross-validation for fine-tuning and testing on the target site. BrainADNet, which does not require pretraining, is trained and tested directly on each target site.

Comparison with classical machine learning, deep learning and generative baselines

Firstly, we present a detailed comparative analysis of our proposed framework against state-of-the-art machine learning and graph learning approaches, demonstrating its effectiveness. Additionally, we analyze the impact of using different brain atlases, namely AAL (116 Regions), Craddock (200 Regions), and Dosenbach (160 Regions), illustrating their influence on classification performance.

In this study, in addition to performing traditional binary classification, we also conduct multiclass classification to simultaneously distinguish between FEDN, rMDD, and HC. The results presented in Fig. 3 clearly indicate that BrainADNet can differentiate between individuals with single versus multiple episodes, as well as those without any depressive disorder. In addition to multiclass classification, we perform binary classifications (MDD vs. HC, rMDD vs. HC, and FEDN vs. HC) to demonstrate the effectiveness of our model in identifying instances from two groups. The results presented in Table 1 show that BrainADNet consistently outperforms all state-of-the-art methods in binary classification as well. Here, we observe that in classifying rMDD versus HC, although our model demonstrates superior accuracy and AUC, it achieves a slightly lower F1 score compared to BrainNetCNN. It is likely due to the significant class imbalance, where HC participants outnumber rMDD patients by nearly a 3:1 ratio. This highlights the need for developing targeted strategies that augment brain signals proportionally to the number of samples in each class. However, to assess potential model bias toward the minority class, we evaluate precision and recall and compare the results against all baseline methods.

Fig. 3. Performance comparison of BrainADNet with state-of-the-art models for multiclass classification.

The classification task involves identifying individuals experiencing their first episode of depression without prior medication; individuals with multiple episodes of depression, and healthy controls within a single classification setting. For comprehensive evaluation, three key metrics are reported: Accuracy, F1 score, and AUC. Classification results obtained using tenfold cross-validation (n = 10). Within each box, the central line indicates the median, the box boundaries represent the interquartile range (IQR), and the whiskers extend to 1.5 times the IQR.

Table 1.

Comparative analysis of our proposed method against state-of-the-art approaches for binary classification

Classification of rMDD vs. HC
		Atlas: AAL		Atlas: Craddock			Atlas: Dosenbach
Classifier	Accuracy (%)	AUC (%)	F1 Score (%)	Accuracy (%)	AUC (%)	F1 Score(%)	Accuracy (%)	AUC (%)	F1 Score(%)
SVM	68.00 ± 6.30	64.40 ± 6.29	38.90 ± 11.09	65.30 ± 6.82	61.90 ± 7.88	43.20 ± 11.15	67.60 ± 4.15	48.70 ± 9.86	15.00 ± 11.56
LR	70.90 ± 5.11	65.30 ± 6.67	41.80 ± 12.31	66.70 ± 7.26	63.40 ± 6.45	40.10 ± 9.61	68.30 ± 2.92	55.60 ± 9.69	40.10 ± 9.29
RF	71.20 ± 2.39	65.20 ± 4.68	27.80 ± 8.58	71.90 ± 4.75	65.60 ± 8.45	31.70 ± 13.72	66.00 ± 3.86	56.30 ± 9.08	15.80 ± 10.46
GCN51	70.74 ± 5.46	68.86 ± 8.83	39.26 ± 20.52	71.75 ± 3.35	72.05 ± 5.76	43.19 ± 10.78	70.26 ± 4.56	69.19 ± 8.34	35.31 ± 19.34
GAT52	69.58 ± 4.30	58.59 ± 6.77	28.07 ± 18.85	70.45 ± 5.48	66.62 ± 8.93	39.13 ± 19.61	68.63 ± 4.15	64.96 ± 6.12	33.54 ± 19.96
GIN53	69.91 ± 4.36	65.65 ± 8.33	29.71 ± 17.21	74.02 ± 3.75	72.47 ± 5.14	45.29 ± 15.23	66.81 ± 5.00	65.04 ± 6.86	29.24 ± 15.15
BrainNetCNN56	72.04 ± 5.16	59.84 ± 7.80	60.46 ± 7.62	72.72 ± 3.33	64.79 ± 6.82	61.70 ± 4.67	67.97 ± 3.78	53.90 ± 4.14	55.17 ± 4.40
TGCN54	68.25 ± 4.35	67.88 ± 6.22	45.71 ± 14.01	67.17 ± 2.74	72.52 ± 4.79	55.89 ± 4.73	61.43 ± 6.29	67.98 ± 4.90	50.89 ± 5.01
BrainGNN55	67.32 ± 4.06	53.34 ± 4.33	22.26 ± 13.47	67.36 ± 3.79	56.41 ± 5.23	33.33 ± 12.21	69.11 ± 6.80	58.19 ± 6.93	35.65 ± 15.37
VAE-GAN60	70.88 ± 4.39	69.51 ± 5.75	43.61 ± 17.98	70.13 ± 5.19	69.51 ± 6.21	43.61 ± 9.23	71.33 ± 4.22	68.00 ± 9.77	35.14 ± 22.00
VAE59	68.60 ± 4.21	67.02 ± 5.37	24.68 ± 16.55	74.28 ± 4.08	68.28 ± 8.20	37.14 ± 15.18	69.78 ± 2.54	66.25 ± 8.70	27.49 ± 19.51
BNT58	74.50 ± 3.11	75.33 ± 3.27	60.18 ± 3.30	75.65 ± 2.53	77.98 ± 2.46	61.18 ± 4.23	71.33 ± 1.63	60.47 ± 2.43	55.06 ± 2.41
PopGCN57	74.87 ± 2.51	70.67 ± 8.88	60.01 ± 8.94	74.28 ± 3.96	67.96 ± 8.16	58.30 ± 12.32	71.33 ± 2.28	60.47 ± 9.16	53.23 ± 10.42
BrainADNet	76.60 ± 5.20	80.10 ± 6.46	58.44 ± 12.13	77.27 ± 3.93	79.80 ± 6.63	57.56 ± 9.20	76.50 ± 5.47	77.98 ± 6.40	53.06 ± 13.92

Classification of MDD vs. HC
	Atlas: AAL			Atlas: Craddock			Atlas: Dosenbach
Classifier	Accuracy (%)	AUC (%)	F1 Score (%)	Accuracy (%)	AUC (%)	F1 Score(%)	Accuracy (%)	AUC (%)	F1 Score(%)
SVM	55.61 ± 2.76	55.13 ± 3.67	58.97 ± 2.50	53.89 ± 3.81	56.24 ± 3.66	58.10 ± 2.94	57.01 ± 3.33	54.76 ± 4.01	62.56 ± 3.19
LR	53.99 ± 3.59	55.21 ± 3.65	58.11 ± 4.03	55.96 ± 1.60	57.12 ± 2.80	59.40 ± 2.37	56.14 ± 3.63	53.90 ± 3.57	61.50 ± 3.63
RF	55.20 ± 4.10	54.96 ± 4.25	55.89 ± 5.15	56.60 ± 3.84	58.45 ± 4.57	60.20 ± 3.63	55.12 ± 3.40	56.35 ± 2.83	58.21 ± 3.78
GCN51	60.51 ± 3.08	65.50 ± 3.40	64.13 ± 3.89	62.10 ± 4.60	67.21 ± 4.15	64.12 ± 4.07	59.35 ± 1.75	59.10 ± 3.03	61.15 ± 4.34
GAT52	59.10 ± 3.74	64.57 ± 3.55	65.23 ± 3.35	59.23 ± 4.34	64.44 ± 5.00	64.23 ± 3.80	57.89 ± 4.99	60.23 ± 4.99	60.10 ± 5.48
GIN53	61.98 ± 2.97	62.56 ± 2.53	62.76 ± 5.44	61.18 ± 3.19	63.34 ± 3.26	61.14 ± 6.30	56.67 ± 4.02	59.21 ± 4.75	56.23 ± 6.92
BrainNetCNN56	59.91 ± 2.84	59.71 ± 2.88	60.12 ± 2.85	60.32 ± 2.24	61.01 ± 2.09	59.12 ± 2.22	57.21 ± 3.78	57.22 ± 3.66	56.78 ± 4.08
TGCN54	57.85 ± 2.06	62.99 ± 3.02	67.50 ± 1.81	58.81 ± 1.63	66.52 ± 2.97	67.10 ± 0.93	55.89 ± 0.98	61.85 ± 2.66	47.82 ± 1.00
BrainGNN55	54.98 ± 2.73	54.34 ± 2.87	57.10 ± 3.12	57.12 ± 4.06	57.11 ± 4.11	58.90 ± 4.49	55.21 ± 3.53	54.95 ± 3.52	60.71 ± 4.68
VAE-GAN60	60.82 ± 2.91	66.07 ± 2.97	60.05 ± 5.12	59.95 ± 2.57	66.44 ± 3.26	63.54 ± 2.84	58.71 ± 3.68	61.88 ± 4.93	61.69 ± 7.16
VAE59	60.45 ± 2.68	65.59 ± 2.91	60.10 ± 6.79	61.39 ± 2.72	65.96 ± 3.66	65.10 ± 2.83	58.23 ± 4.69	62.62 ± 4.74	59.63 ± 8.82
BNT58	60.25 ± 1.56	65.28 ± 1.72	59.93 ± 1.42	66.48 ± 2.16	67.52 ± 1.09	65.51 ± 2.02	64.54 ± 3.52	65.54 ± 3.25	63.54 ± 3.61
PopGCN57	64.92 ± 6.61	65.53 ± 9.42	63.74 ± 8.18	64.23 ± 3.57	66.19 ± 3.48	67.51 ± 3.63	64.54 ± 4.33	64.95 ± 5.70	62.10 ± 4.69
BrainADNet	67.05 ± 2.13	73.59 ± 2.60	69.53 ± 3.13	68.21 ± 2.18	73.40 ± 2.78	68.58 ± 4.54	67.68 ± 4.46	70.26 ± 5.03	64.50 ± 7.20

Classification of FEDN vs. HC
	Atlas: AAL			Atlas: Craddock			Atlas: Dosenbach
Classifier	Accuracy (%)	AUC (%)	F1 Score (%)	Accuracy (%)	AUC (%)	F1 Score(%)	Accuracy (%)	AUC (%)	F1 Score(%)
SVM	58.50 ± 4.90	56.70 ± 6.65	34.20 ± 7.15	55.20 ± 3.81	54.00 ± 6.85	41.10 ± 5.70	57.70 ± 3.64	53.30 ± 6.03	18.70 ± 8.78
LR	59.30 ± 6.84	57.60 ± 7.61	35.40 ± 12.14	58.10 ± 5.43	54.70 ± 7.15	39.00 ± 9.86	57.40 ± 3.57	55.00 ± 4.86	18.00 ± 9.02
RF	63.40 ± 6.05	57.10 ± 5.25	27.40 ± 12.81	64.40 ± 3.88	59.60 ± 5.62	26.80 ± 10.92	55.30 ± 4.25	55.00 ± 5.59	37.80 ± 13.23
GCN51	63.88 ± 6.83	65.17 ± 5.03	39.18 ± 8.78	65.65 ± 2.35	66.05 ± 5.77	41.38 ± 16.66	63.58 ± 2.27	56.96 ± 6.45	15.57 ± 19.81
GAT52	61.20 ± 4.86	59.41 ± 5.42	41.08 ± 9.71	61.47 ± 4.00	58.12 ± 7.15	34.64 ± 18.11	61.15 ± 4.74	59.91 ± 8.33	29.78 ± 21.85
GIN53	63.38 ± 5.36	62.15 ± 7.83	39.78 ± 13.84	63.42 ± 6.45	61.96 ± 6.49	35.16 ± 16.85	61.96 ± 2.91	57.24 ± 5.00	16.82 ± 19.73
BrainNetCNN56	63.72 ± 4.02	55.32 ± 5.65	57.81 ± 7.14	65.16 ± 3.26	56.08 ± 5.90	57.71 ± 7.40	60.32 ± 4.85	51.13 ± 5.30	53.72 ± 6.17
TGCN54	60.47 ± 4.45	65.81 ± 4.50	56.11 ± 4.32	61.21 ± 2.65	68.27 ± 3.66	57.68 ± 3.26	56.24 ± 3.47	64.60 ± 6.58	53.08 ± 7.14
BrainGNN55	60.37 ± 5.69	52.69 ± 6.95	27.65 ± 17.76	58.60 ± 4.38	50.82 ± 5.18	24.27 ± 16.43	57.89 ± 4.66	50.71 ± 4.62	26.48 ± 14.56
VAE-GAN60	66.90 ± 4.68	66.63 ± 5.36	38.60 ± 16.83	64.39 ± 6.09	67.52 ± 8.19	43.74 ± 13.75	60.99 ± 5.23	60.43 ± 5.89	29.83 ± 10.22
VAE59	68.06 ± 4.28	64.80 ± 8.91	46.49 ± 17.16	66.97 ± 8.16	66.37 ± 9.15	43.14 ± 19.03	60.99 ± 3.92	60.43 ± 4.81	29.83 ± 15.71
BNT58	71.22 ± 3.37	76.69 ± 2.06	58.02 ± 3.09	67.74 ± 3.87	72.66 ± 5.99	54.88 ± 12.06	56.97 ± 5.07	62.75 ± 4.71	56.97 ± 2.27
PopGCN57	69.27 ± 2.14	63.54 ± 6.69	57.76 ± 4.38	69.90 ± 3.17	69.97 ± 6.84	59.75 ± 6.32	70.14 ± 5.27	69.18 ± 5.01	62.31 ± 10.31
BrainADNet	73.21 ± 5.87	78.27 ± 5.68	59.51 ± 7.98	72.80 ± 6.07	76.11 ± 6.07	61.65 ± 8.90	73.22 ± 6.15	78.97 ± 7.40	65.06 ± 8.41

Results are reported as mean ± standard deviation of classification performance across experiments, with the best-performing results highlighted in bold.

Comparison with SSL baselines

Here, we compare the classification performance of our proposed method with two state-of-the-art SSL models, namely GATE⁶¹ and GCDA⁴⁸. Since SSL approaches require a separate pretraining phase before fine-tuning and evaluation, we report their performance separately. As shown in Fig. 4, BrainADNet achieves superior performance compared to recent SSL models in all three classification settings, demonstrating its effectiveness in brain disorder classification.

Fig. 4. Performance comparison of BrainADNet and state-of-the-art self-supervised models across three diagnostic settings.

a rMDD vs. HCs, b MDD vs. HCs, and c FEDN vs. HCs. Classification results obtained using 5-fold cross-validation (n = 5). Within each box, the central line indicates the median, the box boundaries represent the interquartile range (IQR), and the whiskers extend to 1.5 times the IQR.

Precision-recall analysis for class imbalance

Class imbalance is a well-recognized challenge in medical imaging problems, as it can bias models toward minority classes and obscure their true predictive performance. The precision-recall (PR) curve captures the trade-off between positive predictive value (precision) and sensitivity (recall), making it particularly well-suited for evaluating potential model bias. The area under the precision-recall curve (AUPRC) summarizes the trade-off between precision and recall across all decision thresholds. Higher AUPRC values indicate a better balance between these two metrics, reflecting more reliable detection of the minority class. Figure 5 presents the PR comparative analysis for three binary classification tasks: rMDD vs. HC, MDD vs. HC, and FEDN vs. HC. The results demonstrate that BrainADNet consistently maintains favorable precision-recall trade-offs in both balanced (MDD vs. HC, FEDN vs. HC) and imbalanced settings (rMDD vs. HC). It achieves superior AUPRC values across all tasks, highlighting its competitive and reliable performance.

Fig. 5. Comparative analysis of precision-recall curves for all baseline models across three diagnostic settings.

a rMDD vs. HCs, b MDD vs. HCs, and c FEDN vs. HCs. The curves characterize the precision-recall trade-off of each model under varying decision thresholds, providing a robust measure of performance in class-imbalanced conditions.

Institutional generalization

We evaluate institutional generalization using the leave-one-site-out (LOSO) cross-validation setting for binary classification between MDD and HC subjects. Given that the dataset comprises fMRI scans from 16 different acquisition sites, we train the model on data from 15 sites and evaluate it on the held-out 16th site. This process is iteratively repeated until each site has served as the test set once. We report the average performance across all folds in the below Fig. 6a. The results demonstrate that BrainADNet consistently maintains strong predictive scores across previously unseen sites, underscoring its robustness to scanner- and site-related variability. This site-level evaluation is particularly important in multi-center neuroimaging research, where heterogeneity in acquisition protocols, hardware, and demographic distributions often hampers model transferability. The consistent performance of BrainADNet under the LOSO setting suggests that the model learns domain-invariant representations, making it well-suited for deployment in real-world, multi-institution clinical applications.

Fig. 6. Comprehensive evaluation of BrainADNet across multiple sites, brain network representations, and demographic groups.

a Comparative analysis of institutional generalization performance for classifying individuals with MDD. Classification performance is obtained using a leave-one-site-out cross-validation strategy across 16 independent sites (n = 16). b Comparison of BrainADNet performance using temporal (time vs. time) and spatial (ROI vs. ROI) brain connectivity networks for classifying MDD versus HCs. Here, T represents time points and R represents brain regions, with each graph offering distinct information for analysis. Classification results are obtained using 10-fold cross-validation (n = 10). c Sensitivity analysis of BrainADNet across male (M) and female (F) groups by comparing the classification accuracy for each gender in identifying MDD individuals. The plot explores potential gender-related biases in model performance when incorporating gender information. Classification results of (b and c) are obtained using tenfold cross-validation (n = 10). For all the boxplots, within each box, the central line indicates the median, the box boundaries represent the interquartile range (IQR), and the whiskers extend to 1.5 times the IQR.

Spatial vs. temporal graph analysis

While our previous experiments focused exclusively on spatial graph representations, prior studies have also explored temporal graphs for modeling brain connectivity. To provide a more comprehensive analysis, we evaluate our framework using both spatial and temporal graphs. Identical sliding-window augmentation and training procedures are applied to both to ensure a fair and consistent comparison. The spatial graph models functional connectivity, where nodes represent brain regions and edges encode statistical dependencies between their BOLD signals, capturing interactions across the entire scanning period. Temporal graphs are generated by taking the transpose of the augmented BOLD signal matrices to compute correlations, capturing similarities between time points and tracking whole-brain activation. As shown in Fig. 6b, spatial graphs consistently outperformed temporal graphs, demonstrating their superior ability to capture clinically informative connectivity patterns for MDD classification.

The observed performance difference likely arises from key neurobiological and modeling considerations. MDD involves persistent disruptions in functional connectivity across large-scale brain networks, especially the default mode network (DMN), salience network (SN), and fronto-limbic circuits^17,62. Spatial graphs are better at modeling these connections, making them more suitable for detecting patterns associated with the disorder. While temporal graphs capture time-varying global activation patterns and may be valuable for studying transient brain dynamics. Moreover, spatial graphs integrate information across the entire available time series, enhancing the signal-to-noise ratio and capturing better connectivity patterns. This analysis highlights the importance of spatial connectivity in understanding brain network alterations relevant to MDD diagnosis.

Sensitivity towards gender

Substantial evidence highlights gender differences in the occurrence of depression, with research consistently indicating that women are more susceptible to depression than men^63–65. Given this established fact, it is essential to evaluate whether our model exhibits any bias toward a specific gender when incorporating gender information alongside fMRI data. Specifically, we assess the ability of our model to accurately classify male and female individuals separately while ensuring balanced performance across both groups. Fig. 6c illustrates the classification performance of our model for males and females separately, along with its overall performance across all data in classifying MDD individuals. Here, Accuracy represents the proportion of all depressed individuals (both male and female) correctly classified by the model; Accuracy (M) denotes the percentage of accurately classified male individuals; while accuracy (F) indicates the percentage of correctly identified female individuals. The results demonstrate that our model achieves consistent performance across both genders, with minimal variations in classification metrics. Despite well-established gender differences in depression prevalence, the model exhibits marginal gender-specific performance disparities. Its ability to integrate gender information while maintaining largely balanced classification performance makes it a valuable tool for identifying depression in diverse populations, supporting its potential applicability in clinical settings.

Ablation study

We conduct a series of ablation studies to evaluate the contribution of each individual component, as well as the effect of different configurations within the proposed framework. Below, we discuss the results and their implications in detail.

Ablation study 1: In this analysis, we incrementally add each component to examine its effect on model performance, and the results are presented in Fig. 7a. Ablation Study 2: This study evaluates the effect on model performance when one component is removed at a time for the classification of MDD patients, as shown in Fig. 7b. The results demonstrate that individual components alone are insufficient for optimal performance. Ablation study 3: Since our framework comprises four modules, we further investigate the contribution of multiple components to MDD classification. Specifically, we perform an additional ablation study in which two modules are removed at a time, allowing us to analyze the effect of the remaining two-module interactions. As shown in Fig. 7c, no two-component combination achieves performance comparable to the full BrainADNet model, highlighting the importance of each module and the overall effectiveness of our framework in brain disorder classification.

Fig. 7. Analysis of different components of the BrainADNet framework.

a Ablation Study 1: classification performance obtained by sequentially adding individual components. b Ablation Study 2: classification performance obtained by removing individual components. c Ablation Study 3: classification performance obtained by evaluating combinations of multiple components. For all boxplots in (a–c), MDD classification results are obtained using tenfold cross-validation (n = 10), and within each box, the central line indicates the median, the box boundaries represent the interquartile range (IQR), and the whiskers extend to 1.5 times the IQR. d Training-loss convergence of different models for MDD classification. e The effect of incorporating demographic attributes on the average performance of the GCN classifier to classify MDD individuals. f Correlations among learned representations in the GCN and BrainADNet models for MDD classification. g Effect of sliding window on average performance of the self-supervised model to classify MDD individuals. h Ablation of the decorrelation regularizer in the average performance of BrainADNet for MDD vs. HC classification. i Training-loss trajectories per epoch for BrainADNet and its ablated variants for MDD vs. HC classification.

In Fig. 7a, the baseline model (Model 1) uses a spectral Chebyshev GCN, which serves as the core classifier. In Model 2, we add skip connections with a dynamic weight fusion block. Skip connections enhance training stability by enabling effective reuse of lower-layer features, which empirically results in smoother convergence and more stable optimization, as shown in Fig. 7d. The fusion block learns to assign adaptive importance weights to outputs from different layers, allowing the model to emphasize discriminative representations while mitigating overfitting through selective feature aggregation. In Model 3, we incorporate demographic information to facilitate more personalized and context-aware representations. Even when added to a vanilla GCN, these attributes lead to noticeable improvements in classification accuracy, as shown in Fig. 7e. This underscores the importance of these features and demonstrates that incorporating demographic information enhances the representational capacity of the model and contributes to improved diagnostic performance. In Model 4, we use a decorrelation regularizer to promote diversity among the learned representations. To assess its effect, we analyze the degree of correlation in the representations learned by a vanilla GCN, as illustrated in Fig. 7f. In the absence of this regularizer, the model tends to produce highly similar features. Although its impact on convergence is modest, the regularizer effectively increases feature diversity, which is essential for improving generalization. Finally, Model 5 (i.e., BrainADNet) augments the training data using a sliding-window approach, generating multiple overlapping time segments per subject. This increases the number and variability of training samples, enabling the model to generalize across dynamic fluctuations in the BOLD signal. In Fig. 7g, we further assess the effect of the augmentation strategy on self-supervised models. Specifically, the GCDA⁴⁸ model is pretrained under two conditions: (i) using the original pretraining dataset and (ii) using the same dataset after applying the sliding-window augmentation. After completing pretraining in both settings, we fine-tune each model on the target site and evaluate its downstream performance. The results presented in Fig. 7a, g demonstrate that employing a proper augmentation strategy is model-agnostic, yielding consistent performance improvements across both supervised and self-supervised learning frameworks. Collectively, the ablation study shows that each module contributes uniquely and complementarily to overall model performance by enhancing convergence, diversity, personalization, and temporal generalization.

Ablation study with decorrelation objective

To empirically evaluate the advantage of decorrelation loss over conventional regularizers, we perform a systematic ablation study. Several variants of BrainADNet are trained to classify MDD patients and HCs, with one or more regularizers (dropout, weight decay, and decorrelation) selectively removed. As shown in Fig. 7h, removing only the decorrelation loss results in an average accuracy drop of ~2%. The best performance is achieved when decorrelation is combined with dropout and weight decay, highlighting their complementary roles in improving generalization and robustness.

Training stability and convergence with decorrelation objective

To assess optimization behavior, we perform controlled ablation experiments where each regularizer (dropout, weight decay, decorrelation) is selectively removed. As shown in Fig. 7i, the model trained with the decorrelation loss exhibits faster and smoother convergence, indicating improved training stability. BrainADNet, which integrates all three regularizers, converges even more quickly and smoothly, highlighting the potential effectiveness of the proposed method.

Discussion

This study aims to advance mental health assessment by improving disease prediction for individuals across different stages of depressive disorder. To this end, we propose an efficient graph-based machine learning framework designed to perform reliably even under limited-data conditions. Our comparative experiments demonstrate consistently strong performance, and a comprehensive ablation study highlights the contribution of each model component to the final prediction. We also examine several practical considerations relevant to real-world applicability, ranging from model performance across different demographic groups to its generalization across diverse clinical and acquisition settings. Finally, we investigate whether depressive disorder affects different demographic groups differently and how functional connectivity patterns evolve across clinical stages. All analyses are conducted using the AAL atlas, and the extended analyses supporting these findings are presented below.

Firstly, we analyze the responsible brain areas in male and female individuals suffering from depressive disorder. Understanding the root cause and the affected regions is essential for the proper treatment of any disorder. Additionally, when treating individuals from different demographic groups, it is critical to consider the unique physiological and neurological differences, as these factors influence both the manifestation of disorders and the effectiveness of treatments. For instance, fundamental differences between males and females, from blood count to brain function, necessitate customized treatment to ensure optimal results. In this study, we analyze the brain regions affected by depressive disorder in both male and female individuals, highlighting gender-specific patterns that could guide more targeted and effective treatment approaches.

In this study, we use gradient weighted class activation mapping (Grad-CAM) with Skip-GCN, allowing localization of the brain areas exerting the greatest influence on the classification performance. Grad-CAM generates importance scores that quantitatively assess the contribution of specific characteristics to predictive accuracy of the model⁶⁶. Grad-CAM was originally developed for convolutional neural networks (CNNs) in the computer-vision domain⁶⁷. Its recent adaptation to GCNs allows the identification of discriminative nodes in non-Euclidean graph structures beyond image space⁶⁸. In this work, we follow the Grad-CAM strategy described in ref. ⁵⁴ to highlight brain regions that uniquely contribute to classification outcomes across male and female groups. We compute Grad-CAM at the GCN input layer using a Gradient × Input approach. The input feature matrix is set to require gradients, and a forward pass generates class scores. We select the softmax probability of the MDD class and backpropagate to obtain its gradients with respect to the input. These gradients are then multiplied element-wise with the original input features to produce a Grad-CAM map highlighting influential input positions. Averaging the resulting values across the feature dimension yields a contribution score for each ROI. This procedure is applied separately for models trained on male and female subjects to identify sex-specific discriminative brain regions.

In Fig. 8, we highlight the top ten brain regions in both male and female individuals that played the most significant role in our model for identifying MDD patients. In Fig. 8c, we highlight the ranking of each brain region for identifying MDD in both male and female subjects, illustrating the extent to which each region contributes to MDD detection within each group. From Fig. 8, it can be observed that, although male and female brains process differently, certain brain regions, such as the angular gyrus (Angular_gyrus_L, Angular_gyrus_R), cingulum (Cingulum_Mid_L), and superior medial frontal cortex (Frontal_Sup_Medial_L), are commonly implicated in both groups. As shown in Fig. 8, distinct regions unique to each gender have also been identified. Research highlights the angular gyrus as strongly associated with MDD^69–72. According to studies, in MDD patients, the left angular gyrus often exhibits reduced regional homogeneity (ReHo) and amplitude of low-frequency fluctuations (ALFF), indicating weakened connectivity that correlates with the severity of depressive symptoms. Also, gray matter volume reductions in the left angular gyrus have been observed, with these structural changes associated with the duration and recurrence of depressive episodes⁷¹. A study showed increased ReHo values and enhanced connectivity in the left angular gyrus after electroconvulsive therapy in MDD patients to better memory recall⁷⁰. As highlighted by various studies, the cingulum also plays a significant role in MDD^73–75. Studies show that MDD patients who have attempted suicide exhibit increased ReHo values in the cingulum compared to those who have not, suggesting heightened brain activity in this region among at-risk individuals. Conversely, these same patients show decreased ReHo in the right cingulate gyrus compared to healthy controls, highlighting altered brain function⁷³. While numerous studies have rigorously validated the remaining responsible brain regions associated with MDD patients as identified by our proposed model, there is still much to discover about how these regions may function differently across genders^76–78. These insights are crucial for advancing gender-responsive treatments and improving outcomes for patients with this debilitating disorder.

Fig. 8. Prominent brain regions in male and female individuals with MDD.

Visualization of the top ten brain regions contributing to MDD classification in a female and b male patients. c Ranking of the top ten brain regions for each group based on their contribution to MDD identification.

In our next analysis, we explore how the progression of depressive disorder reshapes brain functional connectivity. Specifically, we characterize stage-dependent alterations in latent-space connectivity patterns in patients with FEDN and rMDD relative to healthy controls. To detect subtle yet biologically meaningful connectivity differences, we leverage Skip-GCN embeddings, denoted as $Z_{\mathrm{fused}}\in {\mathbb{R}}^{N\times d}$, where N is the number of ROIs and d is the embedding dimensionality. These embeddings provide a high-level, abstract representation of the brain connectivity network.

Figure 9a shows the average learned embeddings for each group (HC, FEDN, and rMDD), which display broadly similar ROI-based patterns. However, the between-group variance maps in Fig. 9b reveal localized differences, with significant variation in specific ROI-feature pairs (left: HC vs. FEDN; right: HC vs. rMDD). These findings demonstrate that BrainADNet learns discriminative latent features that capture subtle but consistent disruptions in brain network topology. Such features contribute to the superior classification performance of the proposed framework. To statistically evaluate group-level connectivity alterations, we conduct edge-wise independent two-sample t-tests (p < 0.05) on functional connectivity (FC) matrices derived from the GCN embeddings. For each subject, the embedding-based FC matrix $Cor\in {\mathbb{R}}^{N\times N}$ is computed using pairwise Pearson’s correlations $Co{r}_{ij}=\mathrm{corr}\left(Z_{i,:},Z_{j,:}\right)$, where Z_i,: denotes the d-dimensional embedding vector for ROI i. This approach captures higher-order dependencies within the latent space, enabling more reliable detection of connectivity alterations.

Fig. 9. Brain functional connectivity differences derived from Skip-GCN embeddings for FEDN and rMDD patients compared to healthy controls (HC).

a Mean Skip-GCN embeddings for HC (left), FEDN (center), and rMDD (right), averaged across subjects. b Between-group variance maps for HC vs. FEDN (left) and HC vs. rMDD (right) were computed as ${\sigma }_b^2={\left({\overline{Z}}_A-\overline{Z}\right)}^2{n}_A+{\left({\overline{Z}}_B-\overline{Z}\right)}^2{n}_B$, where $\overline{Z}=\left(n_A{\overline{Z}}_A+n_B{\overline{Z}}_B\right)/\left(n_A+n_B\right)$, and ${\overline{Z}}_A$ and ${\overline{Z}}_B$ denote the mean embeddings for groups A and B, respectively, while n_A and n_B are the number of subjects per group. c, d Statistical comparisons were performed using two-sided Welch’s t-tests on Fisher z-transformed functional connectivity matrices derived from the embeddings (computed as pairwise Pearson correlations between embedding vectors), with α = 0.05 (uncorrected for multiple comparisons). Nodes represent ROIs, edge widths reflect Hedges' g effect size, and edge colors indicate directionality (red: HC > patient; blue: patient > HC). The top 100 significant edges ranked by p value are displayed.

In Fig. 9c, for FEDN patients, the embedding-based FC reveals pronounced alterations compared to HC, particularly involving the thalamus and its interactions with regions implicated in emotion regulation and decision-making (orbitofrontal cortex), multimodal sensory-language integration (supramarginal gyrus), higher-order cognition, motor execution (superior frontal gyrus, precentral gyrus), etc., consistent with prior findings^79,80. Similarly, in rMDD vs. HC (Fig. 9d), we observe disruptions within regions associated with motor control, postural balance, and coordination (vermis, cerebellum), and its interactions with regions involved in language comprehension, working memory, executive functions, learning, and movement coordination (middle frontal gyrus, hippocampus, and cerebellum) etc., consistent with prior findings^81,82. In summary, this analysis enables the detection of functional connectivity disruptions, providing insights into the neural mechanisms underlying different stages of depressive disorder.

Conclusion

In this study, we introduced BrainADNet, a framework designed to address critical challenges in diagnosing depressive disorders across diverse episodes. By employing a sliding-window technique, BrainADNet enriched and standardized neuroimaging data, thereby overcoming limitations posed by the scarcity of datasets. The integration of a decorrelation regularizer ensured robust learning by promoting independent and diverse feature representations, effectively mitigating overfitting. BrainADNet enhanced the quality of learned representations by leveraging skip connections to reuse lower-layer features and integrating them through a learnable fusion mechanism. Our framework also incorporated subject-specific demographic attributes with brain connectivity features during model training, enabling unified representation learning that captured both network-level neurobiological patterns and individual variability. This holistic approach not only improved diagnostic accuracy but also revealed how depression affects men and women differently, addressing a critical gap in mental health research. Furthermore, BrainADNet revealed how functional connectivity in early, non-medicated, and recurrent stages of depression deviated from that of healthy individuals, thereby offering valuable insights into the evolution of functional connectivity networks. By addressing existing limitations in diagnostic methodology and enabling patient-specific analysis, BrainADNet serves as a valuable tool for advancing depression management.

Author contributions

J.B. contributed to the conception, design of the study, and data analysis. M.Y. conducted the experiments. S.K. and T.K.G. supervised the work and provided critical revisions to the manuscript. All authors reviewed and approved the final version of the manuscript. The DIRECT Consortium shared the data.

Peer review

Peer review information

Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work.

Data availability

Detailed demographic and clinical information for the multiclass classification dataset is provided in the Supplementary Data file in the sheet named Dataset_details. Additionally, the p values corresponding to the results in Fig. 9c, d are included in the Supplementary Data file in the sheet named p_values. Access to the resting-state fMRI dataset can be requested by completing the data request form available at http://rfmri.org/REST-meta-MDD³⁸. Users need to download the zipped dataset and the Data User Agreement from the Data File Download section, and submit the signed agreement as instructed on the website to receive the unzipping password and the accompanying phenotypic information. Further details about the dataset and related studies are available in refs. ^17,83.

Code availability

The training of the proposed model was conducted in Python (version: 3.11.7) using the PyTorch package⁸⁴ (version: 2.3.0), with an Intel(R) Xeon(R) Platinum 8360Y @ 2.40 GHz CPU and NVIDIA A6000 GPU. The source code is publicly available at BrainADNet⁸⁵.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-026-01395-y.