Classification of bipolar disorders using the multilayer modularity in dynamic minimum spanning tree from resting state fMRI

Abstract

The diagnosis of bipolar disorders (BD) mainly depends on the clinical history and behavior observation, while only using clinical tools often limits the diagnosis accuracy. The study aimed to create a novel BD diagnosis framework using multilayer modularity in the dynamic minimum spanning tree (MST). We collected 45 un-medicated BD patients and 47 healthy controls (HC). The sliding window approach was utilized to construct dynamic MST via resting-state functional magnetic resonance imaging (fMRI) data. Firstly, we used three null models to explore the effectiveness of multilayer modularity in dynamic MST. Furthermore, the module allegiance exacted from dynamic MST was applied to train a classifier to discriminate BD patients. Finally, we explored the influence of the FC estimator and MST scale on the performance of the model. The findings indicated that multilayer modularity in the dynamic MST was not a random process in the human brain. And the model achieved an accuracy of 83.70% for identifying BD patients. In addition, we found the default mode network, subcortical network (SubC), and attention network played a key role in the classification. These findings suggested that the multilayer modularity in dynamic MST could highlight the difference between HC and BD patients, which opened up a new diagnostic tool for BD patients.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11571-022-09907-x.

Introduction

Bipolar disorder (BD) is characterized by episodes of depression and mania or hypomania, which brings a heavy burden to patients and their families (Merikangas et al. 2007). According to relevant surveys, only 20% of BD patients with depressive episodes were diagnosed in the first year of seeking therapy, and the average delay time from onset to precise diagnosis was about 5–10 years (Baldessarini et al. 2007; Phillips and Kupfer 2013). Importantly, the clinical evaluation of BD is generally carried out through retrospective clinical information, which may lead to underdiagnosis, misdiagnosis, and poorer outcomes (Wang et al. 2020). Thus, to identify BD patients earlier, it is urgent to develop a method to assist BD diagnosis in clinical.

In recent years, neuroimaging technology has been used to analyze the changes of structural and functional in the human brain (Li et al. 2022). Resting-state functional magnetic resonance imaging (fMRI) using blood oxygenation level-dependent (BOLD) signals as neurophysiological indicators can provide a clue to probe functional interaction between different brain regions (Sankar et al. 2021; Shaffer et al. 2018). At present, using the minimum spanning tree (MST) to analyze brain functional connectivity (FC) has attracted increasing attention (Alexander-Bloch et al. 2010; Ciftçi 2011; Shi et al. 2020). The MST is a sub-network of the original weighted network, which connects all nodes without a closed loop. The sum of weights of all edges is the smallest. MST can ignore pseudo connections in the network and provide topology information similar to the original network (Tewarie et al. 2014).

In previous studies, MST has shown the potential of MST in brain disease recognition. For instance, Firat et al. proposed a graphical model based on MST to recognize cognitive processes through fMRI data. The results indicated that the model performance was superior to the multi-voxel pattern analysis (MVPA) method (Firat et al. 2013). Zare et al. (2016) applied MST to construct a framework to facilitate the early diagnosis of infants with the language-learning disorder. Cui et al. also got a better diagnosis performance in the early mild cognitive impairment (MCI) based on MST (Cui et al. 2018). However, these findings were based on an assumption that FC remains stationary throughout the entire duration of fMRI scans (Allen et al. 2014). In fact, brain activity is spontaneous/intrinsic fluctuations, which are even more prominent in the resting state (Makeig et al. 2004; Onton et al. 2006; Delamillieure et al. 2010). The analysis of dynamic FC is also more prominent in resting-state fMRI (Maleki Balajoo et al. 2020; Shehzad et al. 2009; Ghahari et al. 2020). Similarly, the dynamic MST has been proposed to construct machine learning model to diagnose brain disease (Guo et al. 2017a, b). The previous results suggested that the dynamic MST could capture the critical potential functional changes while static MST fails to. So far, there is a scarcity of research on the static or dynamic MST for BD patients.

The organization of the human brain is governed by functional integration and segregation, which is characterized by modular structure (Eickhoff et al. 2018; Tononi et al. 1994). The modular structure supported specialized information processing, complex dynamics, and cost-efficient spatial embedding (Zamani Esfahlani et al. 2021). Hence, exploring the module structure in the brain network can help us understand the potential mechanism of BD patients. Due to the MST discards the connections that can form a loop, which makes it very difficult to estimate the modular structure in MST. Interestingly, relevant studies showed that the tree and treelike networks also have a high value of modularity (Bagrow 2012; Lee et al. 2006). Compared with random trees constructed by null models, the modularity in actual trees was significant (Bagrow 2012). As a result, we inferred that the modularity in MST could apply to investigate functional integration and segregation for BD patients.

According to previous research, the modularity in the static network has also expanded into multilayer modularity in the dynamic network (Mucha et al. 2010b). Inspired by these work, we constructed dynamic MST using resting-state fMRI data for BD patients and HC in the study. Firstly, we used different null models to verify the multilayer modularity in dynamic MST was not a random process. Here, the module allegiance (MA) matrix was applied to characterize the dynamic functional interaction strength between any two brain regions (Bassett and Sporns 2017). The MA matrix gives the probability of two brain regions being assigned to the same functional module. The larger the value in the matrix, the stronger functional interactions between the two brain regions. On this basis, we proposed a classification framework based on the multilayer modularity in dynamic MST for BD diagnosis. The MA matrix was taken as features to train a Support Vector Machines (SVM) classifier. In addition, we also explored the influence of different FC estimators and MST size on performance of the model. To our knowledge, this is the first time to explore multilayer modularity in dynamic MST over the brain network constructed by fMRI data and apply it on BD diagnosis. Finally, the proposed method shows good potential in the scenario of assisting psychiatric attending physicians in BD diagnosis.

Materials and methods

MRI acquisition and preprocessing

We collected the resting-state fMRI data of 50 HC and 48 BD inpatients from the Psychiatry Department of the Affiliated Brain Hospital of Nanjing Medical University, China. The age range is from 18 to 55 years old. All the participants were right-handed and native Han Chinese. According to the criteria of Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Text Revision (DSM-IV-TR), BD patients were diagnosed by at least 2 psychiatric attending physicians via the Mini-International Neuropsychiatric Interview (MINI, Chinese version) (Sheehan et al. 1998). The 17-item Hamilton Depression Rating Scale (HAMD) scores ≥ 17 (Hamilton 1960). In addition, all patients have not taken antidepressants, antipsychotics or mood stabilizers in the past two weeks, and no psychotherapy or physical therapy during the past six months. The study was approved by the Ethics Committee of the Affiliated Brain Hospital of Nanjing Medical University, Nanjing, China. All participants signed written informed consent after a complete written and verbal explanation.

All participants were required to keep their eyes closed and remind them to stay awake without thinking of anything, to keep the head still, and not to fall asleep during the fMRI scans. The scan parameters of functional images: repetition time/echo time (TR/TE) = 3000/40 ms, field of view (FOV) = 240 × 240 mm², flip angle (FA) = 90°, voxel size = 3.75 × 3.75 × 4 mm³, 32 axial slices with thickness/gap = 4 mm, in-plane matrix = 64 × 64. T1 image: TR/TE = 1900/2.48 ms, FOV = 250 × 250 mm ², FA = 9°, axial slices = 176, thickness = 1 mm, matrix size = 256 × 256.

The fMRI images were preprocessed by SPM 12 (https://www.fil.ion.ucl.ac.uk/spm/) and DPABI (Yan et al. 2016) (http://rfmri.org/dpabi) in Matlab 2016a version. Preprocessing steps included removal of the first 6 vol, slice time correction, realignment were applied to the remaining 127 vol. The subjects head motion exceeded maximum translation > 2 mm, angular motion > 2° and Framewise displace (FD) > 0.3 mm were excluded (Zeng et al. 2014; Van Dijk et al. 2012). T1 weight images were transformed into the Montreal Neurological Institute (MNI) space through the segmentation to get a transformation matrix. The functional images were transformed into standard space via the matrix and spatial smoothing (FWHM = 6 mm, Gaussian kernel). Subsequently, band-pass filtering (0.01–0.08 Hz) was performed in the fMRI time series. Furthermore, the linear regression method was used to control the influence of head motion (Friston 24-parameter) and the signals of the whole brain, white matter and cerebrospinal fluid. Finally, 47 HC and 45 BD patients were included.

Construction of dynamic MST

The time series for HC and BD patients were extracted from the preprocessed fMRI images based on the AAL 90 (Anatomical Automatic Labeling) atlas. We utilized the sliding time window method to partition the entire time series into several overlapping sub-time series of each brain region (Shakil et al. 2016). According to previous studies (Doucet et al. 2017; Eguíluz et al. 2005; Shao et al. 2022; Tao et al. 2013), the Pearson correlation coefficients between all pairs of brain regions’ sub-time series in each window were calculated as the FC matrix. Furthermore, the negative values in the FC matrix were set to zero (Doucet et al. 2017). Subsequently, we applied Kruskal’s algorithm to obtain a MST for each window (Kruskal 1956). In detail, the algorithm arranged the weights of all edges in descending order. Then, the MST was constructed from the edge with the largest weight, and added other edges in order. If a ring was formed, we ignored the current edge and continued to add the next edge. Finally, all nodes were connected into an acyclic subnetwork in each window.

In particular, window length is an important parameter to describe dynamic MST. Undersized windows may result in unreliable estimation and spurious variability in dynamic FC, whereas oversize windows may not capture dynamic information in FC (Hutchison et al. 2013; Cai et al. 2018). Here, the minimum window length should be no less than (1/f_min)/TR (Leonardi and Van De Ville 2015). To explore the influence of window length on the framework, the parameter was set from 35 to 60 TR with an interval of 5 TR (Wee et al. 2016), and sliding with a step size of 1 TR.

Multilayer modularity in dynamic MST

First, we evaluated the multilayer modularity in dynamic MST. Mucha et al. (2010a) developed a function to detect modular structure in multilayer networks. Bassett et al. optimized the function, which was defined as follows (Bassett et al. 2013):

$$Q=\frac{1}{2\mu }\sum _{\mathit{ijlr}}\{(A_{\mathit{ijl}}-{\gamma }_l{P}_{\mathit{ijl}}){\delta }_{\mathit{lr}}+{\delta }_{\mathit{ij}}{\omega }_{\mathit{jlr}}\}\delta (g_{\mathit{il}},g_{\mathit{jr}})$$ 1

where A_ijl is the components of the adjacency matrix for layer l, γ_l is the structural resolution parameter of layer l. The element P_ijl gives the components of the corresponding layer l for the optimization null model. The g_il and g_jr give the community assignment of node i in layer l and node j in layer r, respectively. The elements ω_jlr represent the connectivity strength from node j in layer l to node j in layer r, which is an inter-layer coupling parameter. If l = r, the Kronecker delta δ_lr = 1, and otherwise it equals 0, which is the same in δ_ij and δ (g_il, g_jr). The sum of weights of all edges in the network is $\mu =({\sum }_i{A}_{\mathit{ijl}}+{\sum }_r{\omega }_{\mathit{jlr}})/2$. ${\sum }_i\mathit{Aijl}$ is the intra-layer strength of node j in layer l. ${\sum }_r\omega jlr$ is the inter-layer strength of node j in layer l. Here, the γ_l and ω_jlr were set to 1 based on a previous study (Wei et al. 2017).

We further investigated whether the multilayer modularity in dynamic MST was only a random process. Hence, three different null models (connectional null model, node null model, and temporal null model) were used to verify the multilayer modularity in dynamic MST (Bassett et al. 2011). Subsequently, the generalized Louvain algorithm was applied to detect multilayer modularity in real and random dynamic MST (http://netwiki.amath.unc.edu/GenLouvain/GenLouvain) (Jutla et al. 2011). The maximum Q value indicates optimal modularity. Due to the algorithm can lead to a slight community assignment bias in each independent run, we repeated the algorithm 100 times. Subsequently, we compared the average modularity Q value and number of modules in random and real dynamic MST by a one-sample t-test.

In addition, we calculated an MA matrix (90 × 90) for each individual by marking the functional interaction between two regions in a binary matrix in each window. In detail, if two brain regions were assigned to the same community, the corresponding element in the matrix was set to 1, otherwise it was 0. Then, we averaged the matrices obtained from each window.

BD diagnosis

The steps of proposed classification framework include: feature extraction (MA matrix), selection and classification, as shown in Fig. 1.

Fig. 1.

The analysis flow of the proposed classification framework

To be specific, we converted the upper triangle of the MA matrix into a 4005 × 1 dimensional eigenvector for each individual, resulting in 92 × 4005 features for all participants. The feature dimensions were much larger than the sample size, which can distort the performance of the classifier. To improve the generalization performance, we utilized the principal components analysis (PCA) to reduce the influence of irrelevant or redundant features, which was performed by using dimensionality reduction toolbox (http://lvdmaaten.github.io/drtoolbox/) (Van Der Maaten et al. 2009). SVM is very sensitive to classifying small samples data (Peng et al. 2019; Yin and Hou 2016; Vapnik 1999), especially in identifying psychiatry disorders via neuroimaging data. According to relevant surveys, SVM has been widely used in BD detection (Matsuo et al. 2019; Faedda et al. 2016; Achalia et al. 2020). Here, we also used SVM with sigmoid kernel (Cortes and Vapnik 1995) to train classification model by using the first k principal components (PCs). It can better focus on the comparison of different feature designs on the model performance. It was realized by LIBSVM (https://www.csie.ntu.edu.tw/~cjlin/libsvm/) (Chang and Lin 2011).

Limited by the number of samples, leave-one-out-cross-validation (LOOCV) was utilized to evaluate the performance of the model. In addition, we used nested cross‐validation to correct the bias of the feature selection. The outer loop for assessing the model's performance, the inner loop was performed for feature selection (Fig. 1). Here, the default settings were used for SVM model (Hsu et al. 2003). Subsequently, the accuracy (ACC), sensitivity (SEN), specificity (SPE), area under ROC curve (AUC), Youden index, F-score and balanced accuracy (BAC) were applied to evaluate the final performance of the classification model. Furthermore, we applied permutation test (p < 0.05) to estimate the statistical significance of accuracy, which was repeated 1000 times to construct null distribution of the accuracy. The feature weight underlying SVM can determine the PC that has significant discriminative powers. So the Varimax rotations method was used to estimate the weight of each variable. The weight was considered as the contribution of variables to the PC, while the variables with small contributions could be ignored (Klumpp et al. 2018). Finally, we obtained the most discriminative brain regions and networks in BD patients through weight.

It is worth noting that there are many different ways to calculate FC, such as Spearman correlation (Liang et al. 2009), mutual information (Gómez-Verdejo et al. 2012), zero-lag correlation (van den Heuvel et al. 2009), wavelet correlation (Supekar et al. 2008) and imaginary coherence (Brookes et al. 2011). The choice of the FC may affect the performance of the method. Hence, we further explored the impact of different computing ways of FC on the proposed framework.

The effect of MST scale on classification performance

Based on the relevant survey, the topology properties of MST are also sensitive to network size (Braunstein et al. 2007). Hence, the MST scale may have a potential influence on the proposed framework. To do this, the new dynamic MST was constructed using the Craddock-200 atlas generated by spatially constrained spectral clustering method based on fMRI data (Craddock et al. 2012). The window length and step size were the same as the above setting. Subsequently, we repeated the analysis process.

Results

The demographic and clinical characteristics

The demographic and clinical characteristics were listed in Table 1. There were no significant differences in age, gender, education and mean head motion between HC and BD patients.

Table 1.

The demographic and clinical characteristics

Demographic	HC	BD	Statistical	p
Size (n)	47	45	–	–
Age (years)	31.60 ± 8.25	28.60 ± 9.16	t90 = 1.65	0.10a
Gender (male/female)	24/23	15/30	χ2 = 2.96	0.09b
Education (years)	14.30 ± 2.74	13.80 ± 3.24	t90 = 0.79	0.43a
Seasonal characteristics (Y/N)	–	1/44	–	–
Polarity of first episode (D/M)	–	11/34	–	–
Family history of mental disorder (Y/N)	–	17/28	–	–
HAMA scores	–	17.74 ± 7.36	–	–
Number of episodes of depression	–	2.56 ± 1.73	–	–
Number of episodes of hypomania	–	1.94 ± 1.89	–	–
HAMD scores	–	22.92 ± 4.03	–	–
Mean FD	0.12 ± 0.07	0.10 ± 0.04	t90 = 1.27	0.21a

HC healthy control, BD bipolar disorder

Values are mean ± SD

*p < 0.05

^aTwo-sample two-tailed t test

^bPearson chi-square test

The multilayer modularity in dynamic MST

We used three null models to assess the effectiveness of multilayer modularity in dynamic MST. The modularity Q value and the number of modules were calculated. Compared with null models, the above parameters in real dynamic MST showed a significant difference between both HC and BD patients, see Fig. 2.

Fig. 2.

The significant difference in modularity Q value (first line) and the number of modules (second line) between dynamic MST and null models for HC and BD patients. a1–c1, a2–c2: The statistical results of modularity Q value for HC and BD patients, respectively. d1–f1, d2–f2: The statistical results of number of modules for HC and BD patients, respectively. Abbreviations: HC, human control; BD, bipolar disorder. ***p < 0.001, **p < 0.01, *p < 0.05

Classification performance

To explore the influences of window length on BD diagnosis, we performed these manual designs of proposed framework in each window length. After selecting the optimal number of PCs based on the training dataset, the model obtained the best accuracy of 83.70% (permutation test, p = 0.014) when the window length was 35 TR. The results of different length were shown in Table 2 and Fig. 3. Compared with other studies, we found that the classification performance of our model was superior to most other methods, see Table S1.

Table 2.

The classification performance of different window length

Numbers of PC	ACC (%)	SEN (%)	SPE (%)	Window length (TR)
61	83.70	80.85	86.67	35
19	77.0	82.98	70.00	40
11	71.43	68.09	75.00	45
56	72.83	63.83	82.22	50
21	69.57	63.83	75.56	55
20	70.65	72.34	68.89	60

PC principal components, ACC accuracy, SEN sensitivity, SPE specificity

Fig. 3.

The classification performance of dynamic MST with 35 TR

Further, we explored different computing ways of FC on the classification performance of the method. The results showed that the best performance in BD diagnosis can be achieved with Pearson correlation as FC in the framework (see Table S2).

Most discriminative regions and networks

The findings shown that the model with first 61 PCs obtained the best classification performance when window length was 35TR. After calculating the feature weight of each PC, we found that the PC-58 and PC-3 had the largest discriminative ability for BD patients. Hence, we chose the variables with top 1% weights of PC-58 to estimate the most discriminative brain regions and networks (Shao et al. 2019). In the study, the brain was divided into the subcortical network (SubC), default mode network (DMN), somatomotor network (SMN), visual network (VN) and attention network (AN) (Choi et al. 2012). The findings suggested that the dysfunction of DMN, SubC, and AN played an important role in BD diagnosis, see Table S3. The results were shown in Fig. 4.

Fig. 4.

The most discriminative features for BD patients. a The mean feature weight for each PC underlying SVM. b The contribution of each variable for PC-58. c The most discriminative regions and networks. Abbreviations: SMA, sensory-motor areas; AN, attention network; SubC, sub-cortical network; VIS, visual recognition network; DMN, default mode network

The influence of MST scale on classification

We further investigated the influence of MST scale on performance of the model. We trained a new SVM classifier by using the MA matrix from a larger scale dynamic MST based on the Craddock-200 atlas. The model achieved the best accuracy of 79.35% (permutation test, p = 0.002), the sensitivity of 87.23% and the specificity of 71.11% when the window length was 35 TR. In addition, we compared the current results in a statistical way. In detail, we used the tenfold cross-validation and repeated it 1000 times, then compared the classification accuracies from the two atlas with a two-sample t-test. The accuracy of the model with AAL atlas was significantly higher than that of Craddock-200 atlas (p < 0.01). The results were shown in Fig. 5.

Fig. 5.

The classification performance of new dynamic MST. a The accuracy of the new classifier with different window lengths. b The comparison results of classification performance obtained by using the AAL atlas and Craddock-200 atlas with 35 TR. + , mean accuracy. **p < 0.01

Discussion

In the study, we explored the multilayer modularity in dynamic MST and proposed a classification framework for BD patients. The findings suggested that the method was a useful tool for BD diagnosis. Finally, the influences of the computing ways of FC and the scale of MST on the performance of the method were estimated.

The dynamic MST can be regarded as the backbone of the original dynamic FC network, which is a compact description of the dynamic FC network. It includes dynamic information about the nearest brain region and the shortest linkage pattern of the origin FC network at different times. Based on the organization principles of the human brain, we first explored the multilayer modularity in dynamic MST. We found that the multilayer modularity in dynamic MST for human brain was not a random process via different null models. Furthermore, the findings suggested that the functional interaction between different brain regions was constantly decomposed and reconfigured over time (Damoiseaux and Greicius 2009). Referring to previous studies (Ashourvan et al. 2017; Zheng et al. 2018), we used MA to describe the dynamic functional interaction patterns between different regions. To probe into the discrimination ability of multilayer modularity for BD patients, we applied MA matrix as feature to train a classification model. The results demonstrated that the proposed model achieved the best classification performance when the features extracted from dynamic MST with minimum window length. One can infer that the dynamic variability at the short-term pattern can better reflect the primary pathological difference between HC and BD patients. In addition, the different FC estimators were applied to explore the performance of the method in the study. Compared with other computing ways of FC, the SVM classifier has the best diagnostic ability for BD patients when using Pearson correlation as FC. It indicated that the Pearson correlation was more suitable for evaluating the functional interaction between different brain regions in the framework.

Although the MST is insensitive to the threshold, average connectivity strength and network density, the network scale may affect the MST metrics. Hence, we also explored the effect of MST scale on model performance via a larger-scale dynamic MST. The best accuracy of the new classification model was also obtained by using the minimum window length. Furthermore, we found the diagnosis performance of the model constructed by using the AAL atlas was significantly increased than that of the Craddock-200 atlas. We inferred that MST discarded more connections in original brain FC network with the increase of network size, resulting in the weakening of carrying capacity in MST for the original connection information (Zou and Yang 2019; Tewarie et al. 2015). Hence, we encourage researchers to use the AAL atlas rather than Craddock-200 atlas based on our findings.

Furthermore, we found that the DMN, SubC and AN play an important role in distinguishing BD patients and HC. In previous studies, the above three networks in BD patients were dysfunctional. In detail, as an important component of the brain FC network, the dysfunction in DMN was associated with serious pessimism (Ho et al. 2015). The significant differences in FC of DMN between BD patients and HC have been identified (Yang et al. 2021). The SubC was closely related to emotional processing. The disruption of the network may lead to insufficient inhibition of negative emotional information in the cortex (Elliott et al. 2002). Researchers have also signified disrupted inter-connectivity between DMN and SubC for BD patients (Wang et al. 2017). In addition, the AN also contributed a lot to classification. Attention is one of the important cognitive functions, the activities of brain regions within AN in BD patients were reduced (Fleck et al. 2012). Our findings were consistent with these studies, which support the potentiality of multilayer modularity of dynamic MST in exploring the underlying mechanism of BD patients.

On the whole, the multilayer modularity in dynamic MST is a possible marker of brain functional changes in BD patients. Except for fMRI, the proposed method is also applicable to magneto-encephalography (MEG) and electroencephalogram (EEG) data. Certainly, there were some limitations in the study. First, we used the sliding window method to construct dynamic MST, the window length and step size can affect the stability of the proposed framework. Second, due to the small number of samples, we can’t obtain a sufficiently representative classification accuracy. Third, we need to validate the framework on multi-center data to get reliable results.

Conclusions

In the study, we proposed a classification framework for BD diagnosis based on the multilayer modularity from dynamic MST. The findings showed that the multilayer modularity in dynamic MST is a true expression of neural activity in the human brain. In addition, the results indicated that the multilayer modularity in dynamic MST could provide a new diagnostic marker for BD patients. We also found the scale of the network played an important role in the application of MST. In summary, the proposed framework based on dynamic MST provides a new possibility for diagnosing of BD patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

This work was supported by the National Natural Science Foundation of China (81871066); Jiangsu Provincial Key Research and Development Program (BE2018609 and BE2019675); Jiangsu Provincial Medical Innovation Team of the Project of Invigorating Health Care through Science, Technology and Education (CXTDC2016004); Key Project supported by Medical Science and Technology Development Foundation, Jiangsu Commission of Health (K2019011); Key Project supported by Medical Science and Technology Development Foundation, Nanjing Department of Health (YKK16146 and ZKX18034).

Declarations

Conflict of interest

All authors have no potential conflicts of interest.