To validate the effectiveness of the proposed method, we perform two datasets for disease classification, including the publicly available ADNI dataset and the MDD dataset.

The pipeline of pre-processing for MDD dataset is provided by the REST-meta-MDD project, which consists of head motion correction, spatial normalization, non-linear registration, and spatial smoothing, and so on (Yan et al., 2019).

However, since we only have access to the preprocessed MMD dataset, it is difficult to control the preprocessing pipeline. Therefore, for the sake of fairness, we employ a similar pipeline as in the MDD dataset to pre-process the ADNI dataset in our experiments. We use the Data Processing Assistant for Resting-State fMRI (DPABI) toolbox (Yan et al., 2016) to preprocess the fMRI data.

Specifically, we discard the initial 10 volumes to avoid outliers and performed slice-timing correction. After that, the time series of images for each subject are realigned using a six-parameter (rigid body) linear transformation. Then, we co-register individual T1-weighted images to the mean functional image using a 6 degrees-of-freedom linear transformation without re-sampling. The Friston 24-parameter model is utilized to regress out head motion effects. Finally, we map the fMRI data to the 116 brain regions of the AAL template, convert them into time-series signals, and normalize these signals (Friston et al., 2000). For each subject, we can get 137 time points for each region of interest (ROI) of AAL template after above steps.

The framework of our STDAC scheme is shown in Figure 1, which consists of three modules, including spatial augmentation module, temporal augmentation module, and spatial-temporal fusion module. We denote X=[x1;x2; ⋯;xm]∈Rt×m as the matrix of rs-fMRI data, where m is the total number of nodes, t is the number of time points, xi=[xi1,xi2, ⋯,xit]T is the time series of the ith (i = 1, ⋯ , m) node.

As shown in the Figure 3, we compare our proposed STDAC with three methods, including (1) Original, a scheme that does not use data augmentation, (2) SDA, a scheme only with spatial data augmentation module, (3) TDA, a scheme only with temporal data augmentation module, and (4) STDAC, our scheme. Besides, three popular classifiers are employed to test the effectiveness of our scheme, including support vector machine (SVM), random forest (RF), and artificial neural networks (ANN).

Our STDAC scheme is an offline mode, where the training phase is separated from the testing phase. The training phase shown in Figure 1 has been described in detail in Section 2.3. In our experiments, for the spatial augmentation module, we generate four times more training samples based on the spatial prior knowledge. And for the temporal augmentation module, we generate 100 times more training samples based on the time series. After the training phase, the parameters of the classifier in the test phase are inherited from the classifier in the training phase. We then use the learned classifier to classify the test data.

To obtain an unbiased evaluation, all experiments employ a 10-fold cross-validation method. Specifically, the dataset is randomly divided into ten equal-sized subsets, nine of which are selected for training and the remaining 1 is used for testing. We repeat the whole process 10 times and finally average these 10 results. To prevent learning samples generated from test data during model training, which may result in erroneous experimental results, we divide the data into test dataset and training dataset before the data augmentation stage and only did data augment for the training dataset, not for the test dataset. Besides, we use three indicators (i.e., Accuracy, Precision, and Recall) as performance evaluation metrics, which are commonly used in machine learning classifiers.

Table 1 summarizes the results of three methods with different classifiers in the four-class classification task, and Figure 4 plots the corresponding ROC curves. From Table 1 and Figure 4, we have the following interesting observations.

To further verify the performance of our SDTA scheme, we employ a MDD dataset for the eMCI vs. NC classification task. The experimental results are reported in Table 2. It can be observed that our scheme achieves the best performance in the ablation experiments, which indicates the good robustness of our proposed model. It has been confirmed that data augmentation using spatial-temporal information can help boost classification performance for brain diseases. Besides, the effect of our model is weaker in the MDD dataset than that in the ADNI dataset. The one possible reason is the structural changes in the brain are stronger in neurological diseases (e.g., AD) than in psychiatric diseases (e.g., MDD) (Yan et al., 2019). Another reason is that the REST-meta-MDD dataset comes from 24 different sites, where different scanning devices and environments at different sites can lead to accuracy degradation of the classification model.

We compare our model with the latest data augmentation methods in the field of FBN analysis, including the random window-based data augmentation (Zhu et al., 2021), improved SMOTE (Eslami and Saeed, 2019), auto-encoder-based generative model (Ohno, 2020), conditional generative adversarial network (Raja and Kannimuthu, 2023), and GANSO (Salazar et al., 2021), with experimental results reported in Table 3.

Compared with previous data augmentation methods, our STDAC scheme achieves the best performance on the classification task in the ADNI dataset. Different from the random window-based data augmentation, which uses time windows to generate more samples by obtaining random consecutive time series, our scheme of randomly selecting time periods without caring about continuity would be helpful to obtain an average value when computing the Pearson matrix (Barber et al., 2021). Besides, improved SMOTE, auto-encoder-based generative models, conditional generative adversarial network, and GANSO ignore spatial information and temporal information, which limits their classification performance. In addition, our scheme can fit neural networks better and can generate more data samples than previous methods. Therefore, it can reduce over- or under-fitting caused by insufficient data samples in the neural network.

It is an essential step to select the discriminative biomarkers in FBN analysis. As shown in Figures 5, 6, we show the most discriminative connections and ROIs based on the weights in two classifiers (i.e., spatial classifier and temporal classifier). Since the weights within each subject can be different, we integrate them and select the most discriminative features based on the weights. We demonstrate the top 20 discriminant connections in Figure 5, where the color of each arc is randomly assigned for better visualization.

Figure 6 is a visualization based on the AAL template, where the symbol volume of the nodes represents the discriminative power of each brain region. We normalize the weights of functional connections obtained using the model, then add the weights of functional connections to the brain regions to obtain the weights discriminating brain regions (Zhang et al., 2021). We can observe that the discriminative ROIs of the brain regions after spatial information augmentation are more concentrated in the right temporal lobe, which corresponds to the fact that the right brain regions of the mentioned Alzheimer's patients are more likely to accelerate aging (Roe et al., 2021). The weights of each brain region after temporal information augmentation are relatively more concentrated in the frontal lobes, similar to previous findings (Agosta et al., 2012).

In the proposed STDAC scheme, two data augmentation modules are employed to generate the samples for classification. However, different degrees of data augmentation will lead to different numbers of samples, which may affect the final performance. Therefore, we validate the accuracy of the two data augmentation modules for the classification task of eMCI vs. NC with different degrees.

The experimental results are reported in Figure 7, in which the darkest red on the figures marks the degrees under the best accuracy conditions. It can be observed that the best accuracy is achieved when the spatial degree is 4 and the temporal degree is 100. In addition, whether the degree of data augmentation is too high or too low affects the final result. For example, when the data augmentation degree is higher than 100, the data augmentation effect will also be weakened. The probable reason is that the excessive degree of data augmentation may lead to many duplicate samples. Therefore, we need to select the optimal parameters according to different classification tasks.

The proposed STDAC scheme takes both temporal and spatial information of rs-fMRI data into account, and mitigates the occurrence of over- or under-fitting by data augmentation. However, there are several limitations in our present study. First, individual differences between different subjects are not fully considered in our scheme, which is an important issue for FBN analysis (Folville et al., 2020; Schabdach et al., 2022) and an important direction for our future improvements. Besides, since multi-modal data is taking an increasingly important place in brain analysis (Jia and Lao, 2022; Zhao et al., 2022), the performance of employing one-modal data is limited. Compared with the widely used multi-modal data model (Yu et al., 2021), how to adapt the scheme to the multi-modal data type is a key point in our future work.