FC describes the degree of correlation between blood-oxygen-level-dependent (BOLD) time series in two different brain regions, which has been demonstrated to be a vital tool for revealing various disease mechanisms (Price et al., 2014). The brain atlas divides the entire brain into multiple regions of interest (ROI) based on functional or structural definitions. The average BOLD time series for all voxels within the ROI coordinate ranges represent the BOLD time series for each ROI. FC is mainly obtained by computing the Pearson correlation coefficient of the BOLD time series between paired ROIs. The calculation formula is as follows:

where X_i represents ith ROI time series, S is the standard deviation, cov is the covariance function, and M denotes the number of ROIs in the atlas. Based on the calculated FC value, a symmetric functional connectivity network is established, represented by FCN = [FC_{1, 1}, FC_{1, 2}, ⋯ , FCM,M]∈ℝM×M. The widely utilized atlas is the automatic anatomical labeling (AAL) atlas (Tzourio-Mazoyer et al., 2002; Guo and Zhang, 2020; Ji and Yao, 2021; Pang et al., 2021; Wang et al., 2021, 2022b; Alorf and Khan, 2022; Cai et al., 2022; Hu et al., 2022; Lu et al., 2022; Wang T. et al., 2022; Chen et al., 2023). Besides, FC can be constructed by other atlases, such as the Power atlas (Power et al., 2011; Xu et al., 2020), Craddock 200 atlas (Craddock et al., 2011; Huang et al., 2021; Liang et al., 2021, 2022), Bootstrap Analysis of Stable Clusters (Bellec et al., 2010; Subah et al., 2021; Wang N. et al., 2022), Brainnetome atlas (Fan et al., 2016; Jin et al., 2020), Yeo atlas (Yeo et al., 2011; Gullett et al., 2021), Harvard-Oxford atlas (Desikan et al., 2006; Cao et al., 2020), and Dosenbach atlas (Dosenbach et al., 2010; Zhao et al., 2022). In particular, Zhang et al. (2022) constructed multiple FCNs based on the selected set of the atlas from generated multiple personalized atlases from the AAL atlas to improve the diagnosis effect of MCI.

The above FC refers to the traditional static FC, which reveals the inherent similarity between a pair of ROIs or particular networks (Liu et al., 2021). Dynamic FC is mainly extracted from fMRI data based on the sliding window strategy, which is more helpful than static FC in understanding the basic dynamic characteristics of the human brain (Hutchison et al., 2013). For each subject with M ROIs, an appropriate size and a sliding step of the time window are set, then the time window is slid according to the step over the entire time series. Statistical correlation coefficients such as Pearson are used to calculate the FCN for each window. Continuous sliding and calculation can obtain a set of T FCNs. Dynamic FCN can be denoted as dFCN = [FCN¹, FCN²,⋯ , FCN^T] ∈ ℝ^T×M×M, where T is the number of overlapping windows. Lin et al. (2022) constructed dFCNs based on the sliding window strategy from resting state fMRI (rs-fMRI) data and extracted advanced features of dFCNs to classify brain disease by the proposed convolutional recurrent neural network. Wang B. et al. (2022) constructed dFCNs based on human connectivity project multimodal partitioning. Different from static FCN, dFCN can reveal more useful information for distinguishing between patients with brain diseases and healthy subjects. ElNakieb et al. (2023) proposed an enhanced version of dFCN, where correlation calculations consider temporal dynamics. The accuracy of the presented dFCN achieved 5% higher than traditional FCN. Kam et al. (2020) extracted static brain functional networks (BFNs) from rs-fMRI data by the Group Information Guided Independent Component Analysis method and generated dynamic BFNs based on seed association and sliding window strategy. The experimental results indicated that the simultaneous use of static BFNs and dynamic BFNs prioritizes using static BFNs alone, as dynamic BFNs provide additional time information for BFNs. Li et al. (2020a) obtained better functional brain network estimates based on regulated transfer learning to identify MCI patients from normal controls.

ReHo is a major tool for detecting local neural synchronization (Xing et al., 2021). It reflects the consistency of regional brain activity by computing Kendall's coefficient of concordance between a specific voxel and its neighboring voxels (Zang et al., 2004; Pang et al., 2021). The calculation formula is defined as follows:

where t is the length of the BOLD time series, k is the number of local voxels studied, and r_{i, j} represents the ranking level of a specific voxel j in the entire time series under the ith moment. R_i denotes the rank sum of k voxels at the ith moment. R¯ is the average value at all times. Mean ReHo (mReHo) is obtained by dividing the average ReHo value of the entire brain (Cao et al., 2020; Jia et al., 2021; Pang et al., 2021).

ALFF is an effective fMRI data analysis tool widely used in various diseases to assess resting state activity in brain regions (Zhang et al., 2015; Pang et al., 2021; Zang et al., 2021; Cai et al., 2022). Firstly, the Fourier transform is performed on the time series of each voxel to obtain the frequency domain power spectrum. ALFF is obtained by calculating the square root of the power spectrum and taking the average value over a predefined frequency range (Zang et al., 2007). Mean ALFF (mALFF) is calculated by dividing the mean ALFF value of the entire brain (Cao et al., 2020; Jia et al., 2021; Pang et al., 2021). In addition, several researchers have broadened the range of brain disease classification characteristics from different perspectives. For example, Shi et al. (2021b) extracted intensity-based ALFF histogram features in 2021 to effectively discover regions of the brain with aberrant activity in PD patients. In 2022, Shi et al. (2022) extracted radiological features based on ALFF for the classification study of PD, providing further support for interpreting the PD's pathological mechanism.

As a branch of discrete mathematics, graph theory is frequently employed in network analysis of diverse brain diseases (Farahani et al., 2019; Hallquist and Hillary, 2019; Guo et al., 2020; Yang et al., 2020). The vertices V and edges E are the two most important concepts in constructing the graph G(V, E). Based on graph theory, the ROIs are regarded as vertices V and the connections between ROIs as edges E, thereby constructing a brain network graph for complex brain network analysis.

The metrics in complex brain network analysis mainly include nodal and global measures. Nodal measures characterize the nodal behavior of a network, such as the nodal shortest path length L(i), betweenness centrality BC(i), nodal local efficiency E_loc(i), nodal efficiency E_nodal(i), and nodal clustering coefficient C(i) (Freeman, 1977; Watts and Strogatz, 1998; Latora and Marchiori, 2001; Xu et al., 2020). Global measures indicators reflect the overall network organization, such as clustering coefficient C_P, characteristic path length L_P, small-worldness σ, local efficiency E_loc, and global efficiency E_glob (Watts and Strogatz, 1998; Latora and Marchiori, 2001; Rubinov and Sporns, 2010). Table 2 summarizes the definitions of these nodal and global measures.

Some researchers diagnose disease by comprehensive classification features, integrating more information from various features. For example, Pang et al. (2021) extracted indicators such as ReHo, ALFF, and FC as classification features to provide more comprehensive information on brain function change. Chen et al. (2023) constructed graph measures and FC from rs-fMRI data based on the AAL atlas to distinguish the Idiopathic PD (IPD) and the parkinsonian variant of multiple system atrophy (MSA-P). Jia et al. (2021) extracted mReHo and mALFF from the Brainnetome atlas to distinguish MCI and AD, achieving an accuracy of 86.4%. Ghafoori and Shalbaf (2022) utilized clinical features and functional images extracted from rs-fMRI data to distinguish stable MCI and progressive MCI, with an accuracy improvement of 1.9% compared to using only functional images.

In recent years, multimodal analysis has also been widely adopted in the research of brain diseases by fusing the features of imaging data such as fMRI, structural MRI (sMRI), and DTI to identify more consistent biomarkers and explore the relationship between different modes (Liu et al., 2020b; Zhang et al., 2021; Khatri and Kwon, 2022). Gullett et al. (2021) achieved 94.5% accuracy in classifying stable amnestic MCI (aMCI) and declining aMCI by combining rs-fMRI and T1 sMRI modalities, which is better than only one modality. Shi et al. distinguished SCZ and healthy control (HC) by integrating multimodal information, including fMRI and sMRI. The results revealed that the most discriminate regions between SCZs and HCs mainly locate in the 15 regions, such as the left superior parietal lobule (Shi et al., 2021a). Pang et al. (2022) distinguished IPD and MSA-P by extracting iron-radiomic features, function, diffusion, and volumetric measures from SWI, rs-fMRI, DTI, and T1-weighted imaging, which achieves an accuracy of 91.1%. The classification accuracy was improved to 93.4% by combing clinical variables into the multimodal model. Cao et al. (2020) established a PD diagnosis framework by incorporating relevant clinical features, whole-brain functional connectivity and activity, and gray matter structure, which achieved 84.8% accuracy. The results found that unique whole-brain functional activity and connectivity in PD are mainly located in five networks, such as the affective network. These studies indicated the potential of multimodal data fusion in the auxiliary diagnosis of PD patients.

The fusion of fMRI and genetic data is another research method for multimodal analysis. Using imaging and genetic data to identify biomarkers and classify diseases can help reveal the pathogenesis of diseases, which is of great significance for the development of computational medicine (Bi et al., 2020, 2022; Ghosal et al., 2021). Ghosal et al. (2021) extracted Imaging and Genetics features from task-fMRI and single nucleotide polymorphism to identify SCZ patients from 97 subjects. Bi et al. (2020) constructed ROI-gene pairs based on rs-fMRI and genes to classify early MCI (eMCI) and late MCI (lMCI) patients. The discovery of the most discriminating ROIs and risk genes can provide essential bases for pathological research in the development of MCI.

In addition to the commonly used classification features mentioned above, rs-fMRI time series, 4D fMRI data, and 2D images loaded from fMRI data can be directly used as classification features (Li et al., 2020a; Wang et al., 2020; Ahammed et al., 2021). In particular, Lama and Kwon (2021) adopted the Node2vec graph embedding method to convert brain networks into feature vectors as input to the model. Mousa et al. (2022) calculated the intensity time series of each region and extracted the correlation transfer function (CorrTF) matrix by the CorrTF. It is a promising biomarker for early AD recognition, helping clinicians discover and investigate the affected brain regions and their potential associations during AD development.

The two-sample t-test is a filtered feature selection method to compare whether each feature mean of two independent sample sets has significant differences (Liu et al., 2018). Two-sample t-test requires two groups of samples to be independent of each other and subject to normal distribution. The calculation formula that satisfies homoscedasticity is as follows:

where X1¯ and S₁ represent the mean and standard deviation of each feature of the sample set, and n₁ is the capacity of the sample set. The smaller the t value, the higher the significance level of the feature. The features are classified according to the calculated t value, and the features with weak discrimination ability are removed to improve the classifier's performance. Mousa et al. applied a t-test to select discriminative features between normal subjects and AD patients. Then they trained a support vector machine (SVM) with these features, achieving an accuracy of 99.3% (Mousa et al., 2022). Shi et al. (2021a) proposed a model obtaining an accuracy of 83.5%, which determined the different characteristics between the SCZ and health subjects by two-sample t-test and used maximum uncertainty linear discriminate analysis-based classifier.

Recursive feature elimination (RFE) is a wrapper feature selection approach that recursively reduces the size of the original feature set through multiple training rounds on a base model. SVM-RFE calculates the features' contribution to the model's classification performance by training SVM on the original feature set. Then, remove the features with small contributions and repeat the training process until the remaining features match the required number. Zhou et al. (2022) used the RFE to select and rank the features and revealed the pathological mechanism of MCI through the feature contribution ranking. Zang et al. (2021) reduced the dimensionality of features by RFE before the classification.

The least absolute shrinkage and selection operator (LASSO) is an embedded feature selection method, which is a least square method with an L₁ penalty item (Hsu et al., 2008; Pang et al., 2021; Khatri and Kwon, 2022). LASSO calculates model coefficients β by minimizing the following functions:

where n is the total number of samples, y_i is the category label of the ith subject, x_ij is the jth characteristic of the ith subject, β_j represents the regression coefficient of the jth feature, and m is the feature dimension. α is a non-negative regularization parameter that aims to limit the sparsity of the model. When the bigger α is, the more β_j is set to zero to achieve the purpose of feature selection to reduce unimportant features. Shi et al. (2021b) applied a t-test (P < 0.05) and LASSO to determine the characteristics that significantly differed between PD patients and healthy subjects. Cao et al. selected features using the Mann–Whitney U-test, pre-determined thresholds to screen for highly correlated variables, and LASSO. These selected features were then used to distinguish PD from healthy subjects (Cao et al., 2020).

Logistic regression (LR) is a common method to solve the binary classification problem, which judges the category of samples by comparing the relationship between decision boundaries and samples. Its main idea is to assume that the data obey a certain distribution and make parameter estimation according to the maximum likelihood estimation. Chen et al. (2023) combined a t-test with random forest (RF) to screen out the discriminative features and then distinguish the IPD and MSA-P subjects by LR, obtaining 92.31% accuracy.

RF is an ensemble classification model based on the decision tree, which can process high-dimensional data without feature selection. It randomly deletes some features in decision tree training, then selects features with better classification ability through integrated learning to improve the overall performance. In addition, Bi et al. (2020, 2021b, 2022) optimized the traditional random forest and successively proposed the cluster evolutionary RF, genetic-evolutionary RF, and weighted evolutionary RF algorithms based on genetic data and rs-fMRI data. Improved models can dynamically delete irrelevant or redundant sample features to discover the disease-related risk genes and brain regions.

SVM is a powerful supervised machine learning approach aiming to find a decision boundary in the high-dimensional space that can maximize the interval between two categories. SVM has good generalization and robustness, which has advantages in solving small data sets and over-fitting problems in learning. Yang et al. (2020) diagnosed 236 SCZ patients from 386 subjects by SVM, achieving an accuracy of 71.6%. Gullett et al. (2021) applied SVM to separate stable participants from those with the greatest marginal decline at follow-up with an accuracy of 94.5%. Shi et al. (2022) built SVM to distinguish PD patients from normal subjects, obtaining 81.5% accuracy on the initial training set containing 100 subjects. Furthermore, Bi et al. (2021a) proposed the evolutionary weighted SVM ensemble method to investigate the changes from normal subjects to eMCI to lMCI.

The selection of kernel function and kernel parameter is crucial to SVM's performance. Unlike traditional SVM, multiple kernel SVM (MK-SVM) trains SVM by the kernel that fuses multiple kernel functions. It can effectively learn features of multi-mode data and is more flexible than single-kernel SVM. For example, Liu et al. (2020b) combine structural and functional features by MK-SVM to complete the MCI classification task. Xu et al. (2020) adapted MK-SVM to combine multimodal information to effectively distinguish MCI and healthy subjects, and accuracy was achieved at 92.9%.

Besides the classifiers mentioned above, deep learning models have also been extensively adopted in brain disease diagnosis. With powerful learning and analysis capabilities, deep learning models can automatically find the representation with strong recognition and directly learn the optimal strategy from the original data using hierarchical structures of varying complexity. For instance, Kam et al. (2020) presented a new convolutional neural network (CNN) framework sdMB-CNN, which extracts deep embedded characteristics from static and dynamic functional connectivity networks through 3D-CNN. Compared with traditional methods such as SVM, it significantly improves the diagnosis performance by nearly 10%. Wang et al. presented the multi-kernel capsule network considering the brain's anatomical structure. Set the kernel to match the brain's anatomical structure partition size to capture the connection between regions on different scales (Wang T. et al., 2022). Research shows that the capsule network is viable and prospective in diagnosing SCZ. To directly process 4D fMRI data, Li et al. (2020a) designed a C3d-LSTM for AD recognition, which integrated 3D-CNNs and a long short-term memory network to capture the spatial and time-varying information in the data. Huang et al. (2021) designed a graph-based classification model based on the deep belief network, which is more accurate and efficient than other advanced methods. Ahammed et al. (2021) proposed the DarkASDNet model, predicting the binary classification between ASD patients and normal subjects by 3D fMRI data, with a classification accuracy of 94.7%.

Hold-out validation randomly divides the original data set into independent training and test set (Arlot and Celisse, 2010). This method is simple to operate and avoids the overfitting problems caused by traditional training and testing based on the same data. However, different partitioning methods lead to different results, requiring multiple grouping to ensure the stability and persuasiveness of the model's result. Zhou et al. (2022) evaluated the model's classification performance by the hold-out method. They divided 80% of the dataset into the training set and the rest into the testing set. The dataset was randomly divided by the same proportion for evaluation 100 times. The model acquired an average classification accuracy of 89.4%. Wang et al. (2022b) evaluated the model performance by dividing the dataset into a training set, a validation set, and a testing set at the percentage of 40%, 10%, and 50%, respectively. The hold-out validation was repeated 50 times in their study.

Leave-one-out CV (LOOCV) is the most classical exhaustive CV method, which selects one sample from the original data set as the test data, and the remaining are training data (Arlot and Celisse, 2010; Xu et al., 2020; Cai et al., 2022; Zhao et al., 2022). The experimental data are not affected by random factors because almost all samples in each round are applied to train the model. To avoid the problem of overfitting, Cai et al. (2022) adopted the LOOCV method to estimate the model's classification performance on a dataset containing 67 subjects to overcome the problem of limited data. The obtained results of LOOCV are stable and reliable, ensuring the reproducibility of the experimental process. However, it is more excessive computational overhead than other CV methods, suitable for limited sample sets.

K-fold CV was presented by Geisser (1975) as an alternative to the computationally expensive LOOCV, which divides the dataset into k subsets of almost equal size. Each subset of data is tested once, and the rest k-1 subsets are used as training sets to obtain k models. Combine and analyze the classification accuracy of k models to obtain more stable and accurate model performance. The constant k is variable. When k is the total number of samples in the dataset, k-fold CV is equivalent to LOOCV. Under limited samples, k-fold CV is the simplest and most extensively adopted method for evaluating model performance (Kam et al., 2020; Ghosal et al., 2021; Khatri and Kwon, 2022; Lin et al., 2022).

The articles listed in Table 1 mainly come from two fMRI data sources: internal datasets collected by self or cooperative agencies and public data sets. The former typically come from the same location and have the same collection parameters. The latter often comes from data from different locations, inevitably bringing heterogeneity to the dataset regarding functional magnetic resonance scanners, data collection protocols, and participant population differences, making disease research and analysis more challenging (Ghiassian et al., 2016). However, public datasets typically contain more available data than internal datasets, which can accommodate the demands of more academics for large-scale database analysis. Table 3 summarizes the status of papers and related datasets for the top four brain diseases in the past three years, with 81.8% of the articles using public data. Of these, 90.3% of AD/MCI studies and 100% of ABIDE were conducted on public datasets, demonstrating that well-maintained databases can advance research technology.

Figure 2 shows the relationship between sample sizes and corresponding experimental accuracy. Among them, experiments with high accuracy (79.2% of experiments have an accuracy higher than 80%) are concentrated in the case of sample sizes less than 200 (60% of studies). It can be observed from Figure 2 that the overall trend of AD/MCI and SCZ is that accuracy declines with increasing sample size, which raises concerns about the generalization problem of small sample size research.

Figure 3 displays the histogram of the sample sizes of investigated papers, which clearly illustrates the relationship between the overall sample size and the number of studies. The red dashed line denotes the average, while the blue dashed line represents the median. As depicted in Figure 3, the conducted research in the survey mainly focuses on small sample datasets due to the difficulty in sample collection in clinical practice.

Figure 4 shows a summary of statistical information on the overall accuracy of each disease report. The median accuracies of AD/MCI and PD studies exceed 85%. In contrast, the median accuracy of ASD is less than 80%. The median accuracy of SCZ is between 80 and 85.

Figure 5 shows the validation methods used in the papers. Regarding classifier verification methods, k-fold CV is the most commonly used method (accounting for 58.2%, 32/55 articles). Among them, the commonly used methods are LOOCV and 10-fold CV. Due to a lack of subjects, the hold-out scheme ranked second.