2.1

Participants and MRI acquisition

Rs-fMRI data and T1-weighted images of 51 patients with MD and 21 HCs from the OpenNeuro dataset (ds002748) (https://openneuro.org/datasets/ds002748/versions/1.0.0) were used in this study. All participants were rigorously screened: participants had no mental disorders on a neurological or psychiatric level and were not taking psychotropic drugs or drugs that severely affected blood flow. This study has been approved by the review committees of all participating centers. All participants are informed and consenting.

Imaging information was acquired from the OpenNeuro dataset. The fMRI study was carried out in the International Tomography Center, Novosibirsk, using a 3T Ingenia scanner (Philips). Functional images were acquired using the following parameters: T2*-weighted Ssh echo planar, voxel size 2 × 2 × 5 mm, repetition time (TR) 2500ms, echo time (TE) 35ms. The T1-weighted image was obtained by the T1-weighted 3D turbo field echo, voxel size 1 × 1 × 1 mm. The scan lasted about six minutes, and the participants were asked to remain resting, close their eyes, and not think about anything else.

2.2

Preprocessing of rs-fMRI data

Data Processing & Analysis for Brain Imaging (DPABI) version 4.6 software toolbox was carried out for image preprocessing. The specific preprocessing steps are as follows: (1) remove the first 10 time points to exclude information from the unstable state; (2) slice timing correction and realignment to reduce the impact of factors such as head movement and heartbeat; (3) data point scrubbing: remove the points of framewise displacement > 0.2 mm; excluding the subjects with head movement > 2 mm and rotation > 2°; (4) co-registration for the T1 image and echo planar imaging; (5) segmentation of gray matter (GM), cerebrospinal fluid (CSF), and white matter (WM); (6) regression of nuisance covariates (GM, CSF, and WM); (6) spatial normalization; (7) temporal bandpass filter (0.01–0.08 Hz); (8) image smoothing using a Gaussian kernel with a full-width at half maximum (FWHM) of 6 mm × 6 mm × 6 mm. Particularly, FOCA calculation does not require prior smoothing or filtering.

2.3

Multi-domain brain connectomic parameters calculation

The Neuroscience Information Toolbox (NIT version 1.3, https://www.neuro.uestc.edu.cn/NIT.html) toolbox was used to calculate the FOCA and FCD. The value of FOCA is the product of the time correlation coefficient and the space correlation coefficient calculation, it can be calculated as follows: First, calculate the average FOCA value of each voxel in the whole brain. Secondly, the FOCA value of each voxel divided by the average FOCA value can obtain the normalized FOCA value. Then the normalized FOCA can be used for further statistical analysis.

Global FCD is defined as the number of FCs between a given voxel and other voxels in a binary brain network. Global FCD consists of short-range and long-range FCD. SR-FCD represents the sum of all elements in clusters that are adjacent and directly functionally connected. The long-range FCD can be defined as the global FCD minus the SR-FCD.

The Graph Theoretical Network Analysis (GRETNA) toolbox (https://www.nitrc.org/projects/gretna) was used to calculate the DC value. DC value can be calculated by:

where C_D (N_i) is the DC value of the node N_i, g is the number of nodes in a graph, d_ij represents the number of all edges of the node N_i and the other nodes in the graph, and i ≠ j represents that edges between nodes and themselves are not included.

The Anatomical Automatic Labeling (AAL) template was used to segment the whole brain into 116 brain regions and parameter values for each region were extracted.

2.4

Statistical analysis

The demographic data of 72 participants were processed by Statistical Package for the Social Sciences software, version 24.0 (SPSS, https://www.ibm.com/products/spss-statistics). The chi-square test was used for analyzing the gender differences, and two sample t-tests were carried out for the statistical analysis of age and BDI scores between MD and HC groups.

The parameters of FOCA, SR-FCD, LR-FCD, and DC were analyzed by two-sample t-tests with Bonferroni correction (p < 0.05) based on the Matlab platform. In the process of statistical analysis, age, head movement, and other factors were treated as covariables to avoid affecting the statistical results.

2.5

Feature selection and SVM classification

The LIBSVM toolbox was utilized to train the SVM model. The least absolute shrinkage and selection operator (LASSO)[12] method was performed to further select the features based on the results of statistical analysis for SVM model construction. Hyperparameter tuning was conducted using a grid search approach to identify the optimal values for the penalty parameter (C) and kernel parameter (gamma). A range of values for C and gamma were systematically explored to find the combination that provided the best performance. To ensure robust performance estimation, we employed leave-one-out cross-validation (LOOCV). In LOOCV, each sample in the dataset is used once as a validation set while the remaining samples form the training set. The performance of the SVM model was evaluated using accuracy (ACC), area under the curve (AUC), specificity, and sensitivity.

3.1

Demographic and clinical characteristics

Table 1 presents the demographic and clinical characteristics of 51 MD subjects and 21 HC subjects. The two groups were matched in terms of age and gender (p > 0.05) with no significant differences. However, p < 0.001 indicated that the MD group and HC group have significant differences in depression test scores.

Table 1.

Demographic and clinical characteristics of participants.

a

p value: Chi-square test

b

p value: two-sample t-test.

3.2

Statistical analysis and feature selection

As shown in Table 2 and Figure 2 (A), 12 Brain regions, including 6 regions with significant differences in FOCA, 1 region with significant differences in LR-FCD, 2 regions with significant differences in SR-FCD, 2 regions with significant differences in DC, and 1 region with significant differences both in FOCA and LR-FCD were identified between MD and HC group by two-sample t-test.

Figure 2.

Locations of brain regions associated with each selected feature for each binary classification. A. Brain regions after statistical analysis between MD and HC groups; B. Features selected for SVM model to classify MD and HC groups. (FOCA & LR-FCD, both have significant differences in both FOCA and LR-FCD)

Table 2.

Features for classification of MD and HC.

Using the LASSO method, 10 regions (Table 2 and Figure 2 (B)) with significant differences in the four parameters, including 6 regions in FOCA, 1 region in SR-FCD, 1 region in LR-FCD, and 2 regions in DC were extracted as features for the SVM model to classify MD and HC groups.

3.3

Performances of SVM classification

In the classification task of MD and HC, the SVM model exhibited a high ACC of 81.9%, and an AUC of 0.86, demonstrating the effectiveness of the SVM model in distinguishing depression. The receiver operating characteristic (ROC) curve is shown in Figure 3. While evaluating the SVM model, the specificity was 86.2% and the sensitivity was 71.4%, respectively.

Figure 3.

ROC curve of the SVM model for classifying MD and HC groups. (ROC: Receiver operating characteristic; AUC: area under the curve)

4.1

LASSO for feature dimensionality reduction

Using the LASSO method for feature selection offers significant advantages in neuroimaging analysis. LASSO effectively retains important features by shrinking the coefficients of irrelevant or less significant features to zero, thus simplifying the model. This process helps in identifying key brain regions associated with specific cognitive or behavioral functions. Due to the regularization effect in feature selection, LASSO not only prevents overfitting but also enhances the interpretability of the model. This allows researchers to more clearly understand which brain regions play crucial roles in various neural functions.

4.2

The important role of DMN

Of the 10 important brain regions, 6 regions are located in DMN. This suggests that DMN is strongly associated with depression. The DMN is a group of interconnected brain regions that show high activity when a person is at rest, introspecting, or thinking state. Key areas of the DMN include the prefrontal cortex, posterior cingulate cortex, medial prefrontal cortex, and medial temporal lobes[13]. It is associated with self-awareness, memory recall, future planning, and social cognition. Abnormal activity or connectivity in the DMN was reported linked to depression, indicating its crucial role in depressive conditions[14]. Zhang et al. studied the classification of major depressive disorder, subclinical depression, and HC groups. This study reported 48 effective brain regions as features to construct SVM models, among which several functional connectivity features were associated with DMN[10]. Another study[15] based on meta-analysis demonstrated the abnormal connectivity between regions in DMN and frontoparietal.

4.3

Limitations

This study still has limitations. First, the sample size of the dataset is far from enough, which may affect the accuracy of the findings. Extending the data set can enhance the robustness of the results. Secondly, only rs-fMRI data was relatively simple. Integrating other modalities, such as task-based fMRI, and structural MRI can provide a more complete understanding of the underlying neural mechanisms.