Conventional FC-based methods usually first extract handcrafted features from functional connectivity networks and then train a classifier (e.g., support vector machine, SVM) for disease prediction (Bai et al., 2009; Jie et al., 2014a,b; Plis et al., 2014; Bi et al., 2018). For example, Bi et al. (2018) designed a random SVM cluster method for AD identification. This method firstly randomly selected samples and FC features to establish multiple SVMs, and then employed an ensemble strategy for the final prediction. Jie et al. (2014a) extracted and integrated multiple properties of static FC networks (e.g., connectivity strength and local clustering coefficient) for diagnosing brain diseases and achieved better performance compared with single network measures. Even so, these methods usually rely on handcrafted feature representations for classification models, thereby possibly producing sub-optimal classification performance.

More recently, deep learning methods have been proposed to automatically learn data-driven features from dFC networks. These methods offer a unified framework for fMRI feature learning and brain disorder prediction, ultimately achieving better performance. For example, Liang et al. (2021) proposed a novel convolutional neural network combined with a prototype learning (CNNPL) framework to classify brain functional networks for the diagnosis of autism spectrum disorder. Specifically, it used traditional convolutional neural networks to extract high-level features from pre-defined FCs and further designed a prototype learning strategy to automatically learn prototypes of each category for ASD classification. Eslami et al. (2019) proposed to extract the lower dimensional feature representation of FCs using an autoencoder, followed by a single layer perceptron (SLP) for ASD identification. Kawahara et al. (2017) developed three distinct convolutional layers—edge-to-edge (E2E), edge-to-node (E2N), and node-to-graph layer (N2G)—to capture the spatial characteristics of structural brain connectivity for cognitive and motor developmental score prediction in premature infants. Due to the graph-structured nature of brain functional networks, graph neural network (GNN), which can learn expressive graph representations, have shown significant potential in FC-based brain disease diagnosis. For example, Ktena et al. (2018) proposed learning a graph similarity metric using a siamese graph convolutional neural network for ASD classification. Yao et al. (2021) developed a mutual multi-scale triplet graph convolutional network for brain disorder diagnosis using functional or structural connectivity. Li et al. (2020) designed an ROI-aware graph convolutional layer that leveraged fMRI's topological and functional information for ASD diagnosis.

These deep learning methods greatly improve the efficiency and classification/regression performance in FC-based analysis due to their end-to-end architecture. However, these methods mainly study static patterns of brain networks, thereby ignoring the dynamic characteristics of brain FCs. Besides, GNN-based methods generally take a fixed graph structure as input, whose reliability remains to be discussed.

Several dynamic functional analysis methods have recently been proposed for brain disease classification (Wang et al., 2019b, 2023; Yan et al., 2019; Gadgil et al., 2020; Lin et al., 2022; Liu et al., 2022; Liang et al., 2023). For example, Wang et al. (2019b) proposed a spatial-temporal convolutional-recurrent neural network (STNet) for Alzheimer's disease progression prediction using rs-fMRI time series. Specifically, a convolutional component was first employed to construct the FC within each time-series segment. Then, the long short-term memory (LSTM) units were used to model the temporal dynamics patterns of these successive FCs. Finally, a fully connected layer is used to perform disease progression prediction. Lin et al. (2022) developed a convolutional recurrent neural network (CRNN) for dynamic FCs analysis and automated brain disease diagnosis. In this method, a sequence of pre-constructed FC networks was input into three convolutional layers to extract temporal features, and an LSTM layer was used to capture temporal information from multiple time segments, followed by three fully connected layers for brain disease classification. To take advantage of spatio-temporal information of fMRI data, Yan et al. (2019) designed a multi-scale RNN framework to classify schizophrenia. Specifically, stacked convolution layers were used to extract different scale features, followed by a two-layer stacked Gated Recurrent Unit (GRU) to mine dynamic information conveyed in fMRI series. Gadgil et al. (2020) trained a spatio-temporal graph convolutional network (ST-GCN) on each segment of the BOLD time series to predict gender and age. In this method, a positive and symmetric “edge importance” matrix was first integrated to determine the importance of spatial graph edges. Then, three layers of ST-GC units were used to perform spatial graph convolution, followed by a fully connected layer for final prediction. Liang et al. (2023) proposed a self-supervised multi-task learning model for detecting AD progression, in which a masked map auto-encoder and temporal contrast learning were jointly pre-trained to capture the structural and evolutionary features of longitudinal brain networks. Liu et al. (2022) proposed a method based on nested residual convolutional denoising autoencoder (NRCDAE) and convolutional gated recurrent unit (GRU) for ADHD diagnosis. Specifically, the NRCDAE was used to reduce the spatial dimension of rs-fMRI and extract the 3D spatial features. Then, the 3D convolutional GRU was adopted to extract the spatial and temporal features simultaneously for classification. Although existing dynamic FC-based methods consider the temporal dynamics in the prediction of disease progression, those methods fail to capture the global temporal changing patterns of the whole brain (i.e., the longitudinal network-level patterns).

As illustrated in Figure 1, the proposed ASTNet includes (1) partition of rs-fMRI time series, (2) adaptive functional connectivity generation, and (3) temporal dependency mining.

We implement the framework using Python 3.7 and Pytorch library. For each subject, the adjacency matrix is constructed via our designed AGL strategy, where a random initialization technique initializes the vector ω according to a normal distribution. Subsequently, the graph convolution process comprises three GCN layers, followed by batch normalization, ReLU activation, and a dropout rate of 0.5. We then perform a dot product operation on the representation generated from the GCN layers to construct symmetric matrices describing the degree of correlation between nodes. To reduce dimensionality, we flatten the upper triangular portion of each matrix into a vector. This vector is subsequently fed into an MLP consisting of three fully connected layers. Additionally, we incorporate two dropout layers to mitigate overfitting.

Subsequently, we employ an attention mechanism to obtain attention weights and compute the weighted sum of the vector data. This process yields a 32-dimensional vector, which serves as the final output of the global branch. For local analysis using the BiGRU, we perform experiments with different numbers of units, specifically 4, 16, and 64, for training data from various sites.

In the experiments, we compare our ASTNet model with the following eight methods, including the baseline methods and any other variants of the proposed method.

Note that, we both have static and dynamic experiments for method MLP and GCN, named as MLP_s, MLP_d, GCN_s, and GCN_d. And we apply global branch in the TDM to mine temporal dependency for MLP_d and adopt the whole TDM module for GCN_d classification.

We evaluate the proposed method on five different sites (i.e., PKU, NYU, OHSU, NI, and KKI) of the ADHD database based on rs-fMRI data. We divide data on each site into training data and test data, following Global Competition. The test set is unseen during the training stage.

To evaluate classification performance, three metrics are used, including accuracy (ACC), sensitivity (SEN), and specificity (SPE). These metrics are defined as follows: ACC = (TP + TN)/(TP + TN + FP + FN), SEN = TP/(TP + FN), SPE = TN/(TN + FP). Here, TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative values, respectively. Higher values for these metrics indicate better classification performance.

The quantitative results achieved by different methods in the binary classification tasks are reported in Table 2. From Table 2, one could have three main observations.

First, our proposed method and its variants (i.e., AGL_d and GCN_d) generally achieve better performance compared with the baseline methods (i.e., MLP, BiGRU, and Auto-Encoder) in the classification task. For example, in terms of ACC values, ASTNet achieved an improvement of 13.9%, compared with the best baseline method (with 66.6%) in ADHD classification. This demonstrates that our designed adaptive functional connectivity learning strategy and temporal dependency mining module can help extract more discriminative fMRI features, thus enhancing classification performance. Second, our proposed ASTNet and its variants outperform those methods without considering the temporal dynamics (e.g., with GCN_s and AGL_s) in terms of most metrics. In particular, the SEN values produced by our ASTNet in site PKU and NYU are 85.1 and 75.0%, which is higher than other methods. These results suggest that our TDM module can effectively capture dynamic changes in rs-fMRI time series. Finally, our ASTNet is superior to its variants (i.e., AGL_d and GCN_d). This result implies that the adaptive functional connectivity learning strategy and TDM module help boost the learning performance of ASTNet.

The proposed ASTNet can automatically learn dFCs in a data-driven manner, which differs from previous studies that rely on predefined FC networks (e.g., via Pearson's correlation). We now further analyze the FC networks learned by the proposed adaptive method. Specifically, as introduced in Section 3.2.2, the AFCG can learn new connectivity strength between each central node/ROI and all the remaining N-1 ROIs. Therefore, we can generate a fully connected FC network based on the learned connectivity vector. Given the size of the sliding window, different lengths of data will result in different numbers of segments. Taking PKU site as an example, we can construct K = 11 dynamic FC networks for each subject, with each network corresponding to a segment. Finally, using the standard t-test, we measure the group difference between ADHD and NC via p-values, with group difference matrices visualized in Figure 2. For comparison, in Figure 2, we also report the group difference of the stationary FC network that is constructed via measuring Pearson correlation coefficients between fMRI time series of pairwise brain ROIs. Note that the obtained p-values were binarized (i.e., setting p-values more than 0.05 to 1 and 0 otherwise) for clarity in Figure 2. From Figure 2, we have the following observations.

First, from Fⁱ, i = 1, …, 11 in Figure 2, we can observe that the group difference matrices generated by different segments exhibit significant differences, which further validates the temporal variability of brain networks. Second, by comparing our learned Fⁱ and Static in Figure 2, it can be found that the dynamic FC network learned by our ASTNet shows superiority over the pre-defined static network in identifying disease-related functional connectivities and ROIs. For example, Several ROIs, such as the anterior cingulate and paracingulate gyri node (ACG.L) in F¹, the superior parietal gyrus node (SPG.L) in F⁶, and the cerebellum nodes in F⁵, are detected by our dynamic FC networks in AD vs. NC classification. These findings aligns with previous AD-related studies, which further demonstrates the learned FCs by our ASTNet have good interpretability.

We further compare the proposed ASTNet with seven state-of-the-art (SOTA) methods designed for ADHD analysis, including PCA-LDA (Dey et al., 2012), EM-MI (Dou et al., 2020), 3D CNN (Zou et al., 2017), SASNI (Zhang et al., 2017), SPAE (Cao et al., 2023), STAAE (Dong et al., 2020), and KD-Transformer (Zhang et al., 2022). Note that all the methods use the standard training/test sets division by the data set. The classification results achieved by different methods are reported in Table 3, with the best results highlighted in bold. From Table 3, we can have the following findings.

First, the proposed ASTNet outperforms seven SOTA methods in ADHD classification task on the ADHD-200 dataset, which implies that our ASTNet can learn more discriminative features for ADHD identification. Second, Compared with static methods (i.e., PCA-LDA, 3D CNN, SASNI, and KD-Transformer), the methods (i.e., SPAE, STAAE, and KD-Transformer) that consider temporal dynamics in fMRI series achieves relatively better performance. This suggests that temporal information conveyed in fMRI series plays an important role in distinguishing ADHD patients from normal controls.

Third, the ACC of our ASTNet achieves an improvement of 5% compared with the STAAE that designs a spatiotemporal attention auto-encoder long-distance dependency in time. This finding further demonstrates the superiority of our ASTNet in dynamic brain network learning and brain disorder classification.

In main experiments, we divide fMRI series using sliding window strategy with window size of 20. To investigate the influence of different sliding window sizes on results, we vary the values of sliding window size within [10, 15, ⋯ , 30]. The results in ADHD classification on five sites are reported in Figure 3. As shown in Figure 3, we can see that our the classification accuracy of our ASTNet fluctuates to a certain extent as the window size increases. When window size is 20, our method achieves its peak performance across different sites, which validates that our selected window size is reasonable.

To demonstrate the effectiveness of each module in the proposed ASTNet, we further compare our ASTNet with its degenerated variants, including (1) GCN_d without incorporating adaptive graph learning (AGL) module, (2) AGL_d without incorporating GCN module, (3) ASTNet_G without global branch and (4) ASTNet_L without local branch. The experiment results of our ASTNet and its variants are reported in Table 4.

It can be found from that our ASTNet consistently outperforms GCN_d that fails to adaptively learn FC strength. This implies that our designed adaptive graph learning strategy can automatically generate more reliable FC network for subsequent analysis, thus boosting model performance. In addition, we can observe that our ASTNet is superior to ASTNet_G without modeling long-term dependencies among dynamic functional connectivities (dFCs). Besides, our ASTNet achieves better performance than ASTNet_L that can not capture local temporal dependency in dFCs. These observations further demonstrate the advantage of our ASTNet, which simultaneously uses global and local branches in fMRI temporal feature learning.

While our work achieves good results in automatically identifying ADHD using fMRI data, several issues still need to be considered in the future to further improve the performance of the proposed method. First, considering the small-sample-size issue of fMRI data, we will employ transfer learning and pretraining strategies to further enhance model generalization. Second, different brain image modalities, such as structural MRI and Positron Emission Tomography (PET), can provide complementary information for ADHD diagnosis. Integrating multimodal neuroimages would be an interesting avenue to pursue, which will be our future work. Finally, we only construct the functional connectivity matrix based on the AAL atlas with 116 pre-defined ROIs in this work. In the future, we will explore multi-scale functional connectivity networks divided by multiple brain atlas to capture complementary topological information.