Pattern recognition of schizophrenia based on multidimensional spatial feature fusion

Abstract

Purpose

Schizophrenia (SCH) is a severe psychiatric disorder associated with brain connectivity abnormalities, and early diagnosis can significantly reduce the burden on the families of the patients. Though several classification methods have been created to identify SCH, a reliable method is yet to be found. In this study, we explore the performance of multidimensional spatial feature fusion in the recognition of SCH.

Materials and methods

Using an MRI connectomes dataset, we extract the spatial pattern network (SPN) and diffusion map embedding (DME) features from functional connectivity (FC) and structural connectivity (SC) networks of both schizophrenic patients and healthy subjects, and we use both single mode features and fused features to classify the two groups.

Results

Compared to the single mode features, the fused features showed superior performance in classification. By fusing the SPN and DME features of the structural network, we obtained the highest accuracy of 87.50%.

Conclusions

Multidimensional spatial feature fusion is promising as a reliable method for the recognition of SCH.

Keywords:

• Schizophrenia
• brain network
• pattern recognition
• multidimensional spatial feature fusion

1. Introduction

Schizophrenia (SCH) is a typical chronic mental disease characterized by the hallucinations, social withdrawal, and cognitive deficits. It has been reported that about 1% SCH patients suffered lifetime prevalence, which leads to a huge burden on their families [1–3]. This disease can be caused by genetic diseases, environmental factors, stress, and psychology but the specific pathology is unclear. Due to the complexity of the pathology and the symptoms, diagnosis of SCH mainly relies on semiotic diagnostic protocols. The criteria and manuals provide a standardized process and a basis for the diagnosis. However, due to the lack of a clear understanding of the pathology and mechanism of SCH, this current diagnosis method, which is based on the classification of symptoms, is quite unstable in many cases.

With the development of technology, high-field magnetic resonance devices allow researchers to obtain higher-quality brain images and have become a noninvasive and highly effective way to study SCH [4]. Connectivity analysis based on brain images can describe the connectivity network of the brain [5]. In the past few years, many studies have shown that there are changes in the connectivity of some specific brain regions in the brain networks of SCH patients [6–9]. McHugo et al. used these differences in brain connectivity to characterize the pathology of SCH and explored the possibility of using them as novel biomarkers for predicting SCH [10]. As another approach, many researchers try to achieve more stable and reliable recognition through machine learning or deep learning algorithms. For example, Schnack et al. tried to distinguish 1.5 T magnetic resonance images of 46 schizophrenic patients and 43 healthy subjects by a support vector machine (SVM) in 2014, and the recognition accuracy was approximately 76% [11]. In 2020, Oh et al. used deep learning methods to achieve a 71–90% classification effect in five MRI databases of SCH [12]. Wang et al. proposed an innovative method for identifying SCH based on multikernel capsule network, achieving the highest accuracy rate of 82% [13]. Gutiérrez-Gómez et al. improved the classification effect of 27 SCH patients and healthy subjects in the Zenodo platform dataset through a SVM with recursive feature elimination, with an accuracy around 76% [14]. Although the above works achieved the recognition of brain diseases based on implicit differences of the brain image signals, their recognition accuracy is not precise enough to be stably applied to an actual diagnosis.

Compared to direct statistical measurements, the spatial features of brain networks may contain more detailed information and could be meaningful for differentiating various network patterns. As promising approaches, common spatial pattern (CSP) and diffusion map embedding (DME) algorithms have been developed to extract spatial features for distinguishing brain networks under different conditions [10]. Based on a CSP algorithm, the spatial filters that guarantee the maximization of the variance of one class while minimizing that of another class can be obtained. Remarkably, previous studies have demonstrated that spatial pattern network (SPN) features are highly effective in identifying refractory epilepsy without structural abnormalities, as well as in differentiating patients with SCH from healthy controls. The DME algorithm, as a relatively novel approach, extracts the local gradient features of the connectivity matrix by learning the main feature vectors reflecting the distance change, and has been successfully applied to image analysis and image processing.

Many studies have shown that the low-dimensional features in a single mode lack effective discrimination abilities when reflecting the complex brain network related to brain diseases, and feature fusion plays a vital role in pattern recognition by enhancing the discriminating ability and robustness of the recognition system [15]. So, an effective scheme is to achieve a more comprehensive characterization by fusing the heterogeneous information in multimodal data [16–19]. In our study, we extract the SPN and DME spatial features of the functional and structural brain connectivity of both the SCH and the healthy subjects. We adopt both single-mode feature and multidimensional feature fusion strategies to train and test the classifier. Our results indicate that the DME features can be applied to connectivity feature extraction, while the fused DME and SPN features with comprehensive information could further improve the classification performance to differentiate SCH patients and healthy subjects.

2. Methods

2.1. Dataset

The structural and functional MRI connectomes dataset are provided by the 27_SCHZ_CTRL_dataset for SCH differentiation [20] (https://doi.org/10.5281/zenodo.3758534), which includes 27 healthy subjects (age 35 ± 6.8 years) and 27 schizophrenic patients (age 41 ± 9.6). The written consent was obtained for all subjects – in accordance with institutional guidelines of the Ethics Committee of Clinical Research of the Faculty of Biology and Medicine, University of Lausanne, Lausanne, Switzerland. All the subjects in the dataset were scanned in a 3-Tesla MRI scanner using a 32-channel head-coil, and the initial signal of the dataset consisted of MPRAGE, DSI and resting-state-fMRI data. The gray and white matter segmented from the MPRAGE volume were parcellated into 83 cortical and subcortical areas, and furtherly subdivided into 129, 234, 463, and 1015 parcels according to the Lausanne anatomical atlas [21–23] (https://github.com/jvohryzek/bert4lausanne2008). The density, number and length of fibers, the generalized fractional anisotropy (gFA) and the average apparent diffusion coefficient (ADC) were included in the dataset to measure SC. A functional feature refers to the Pearson correlation between the individual pairs of regions.

As a reference, we selected 234 × 234 resolution data points, while gFA was used as a quantitative measure of the structural connectivity (SC) and compared with the functional connectivity (FC, the Pearson correlation) to identify schizophrenic patients. The gFA, as one of the main parameters of DSI, reflects the difference in the diffusion direction of the water molecules within the tissue, which is known to be a marker of tract integrity [24]. FC matrices were constructed by calculating the Pearson correlation between individual brain regions’ time series, which is an index of the correlation between two areas of the brain.

2.2. Common Spatial patterns

The CSP algorithm can extract the potential spatial information in the brain network, which helps distinguish between SCH patients and healthy subjects [25–27]. The principle of CSP is to find a group of spatial filters, under which the variance of one class reaches a maximum while the other turns minimal, making the two classes become easily distinguishable [28,29]. Mathematically, when ${\phi }_1$ and ${\phi }_2$ represent the adjacency matrices of the SCH patients and healthy subjects, respectively, and $\omega$ represents the spatial filter (i.e., the projection), the variance of the two classes under the projection can be expressed as ${\phi }_1\omega$ and ${\phi }_2\omega ,$ where ${\Phi }_1$ and ${\Phi }_2$ represent the spatial covariance of the two classes, so that the CSP transformation can be converted into a Rayleigh quotient maximization problem [8,30,31]: (1) $$J\left(\omega \right)=\frac{{\omega }^T{\phi }_1^T{\phi }_1\omega }{{\omega }^T{\phi }_2^T{\phi }_2\omega }=\frac{{\omega }^T{\Phi }_1\omega }{{\omega }^T{\Phi }_2\omega }.$$(1)

Since the value of $J(\omega )$ does not change with the scaling of $\omega ,$ we will find that the scaling of $\omega$ satisfies the equation ${\omega }^T{\Phi }_2\omega =1.$ By introducing the Lagrange multiplier, the problem in (1)(1) $$J\left(\omega \right)=\frac{{\omega }^T{\phi }_1^T{\phi }_1\omega }{{\omega }^T{\phi }_2^T{\phi }_2\omega }=\frac{{\omega }^T{\Phi }_1\omega }{{\omega }^T{\Phi }_2\omega }.$$(1) can be converted to the following functions: (2) $$L\left(\omega ,\lambda \right)={\omega }^T{\Phi }_1\omega -\lambda \left({\omega }^T{\Phi }_2\omega -1\right).$$(2)

When $\frac{\partial L}{\partial \omega }=0,$ $L(\omega ,\lambda )$ reaches the extreme value, and the objective projection $\omega$ can be estimated using the generalized eigenvalue equation: (3) $${\Phi }_2^{-1}{\Phi }_{1}W=\sum W$$(3)

Here, $W$ is the eigenvector matrix of ${{\Phi }_2^{-1}\Phi }_1,$ $\sum =d\mathit{iag}({\lambda }_1,{\lambda }_2,\dots {\lambda }_m)$ is the diagonal matrix composed of the corresponding eigenvalues, and the above equation transforms the problem into a solution of the ${{\Phi }_2^{-1}\Phi }_1$ eigenvalues. m represents the number of SPN filters, the difference of filters is determined by its eigenvalue, and the spatial feature extracted by the first pair of filters with the largest eigenvalue has the strongest discrimination ability. Specifically, the n pairs of discriminative SPN filters extract 2n features, and for the 234 × 234 adjacency matrix $M,$ each spatial filter extracts a feature vector of 234, while $V_{\mathit{SPN}}$ is a $234\times 2n$ matrix. Finally, the variance of the weighted different brain regions is calculated as follows: (4) $$F_{\mathit{SPN}}=\log \left(\mathit{var}\right(V_{\mathit{SPN}}^{T}M\left)\right)$$(4) where $\mathit{var}(·)$ represents the variance operation of each row of the SPN transformation matrix, and the corresponding SPN feature expressed as a vector of length 2n.

2.3. Diffusion map embedding

Diffusion map embedding is a nonlinear dimensionality reduction method based on manifold learning, which helps to reveal the complex structure of the high-dimensional data of the brain network [8,30–35]; the principle of DME is to construct a diffusion map on the original high-dimensional data, where the distances of the strongly connected brain regions are close, the distances of the weakly connected brain regions are far, the data are reduced, and the low-dimensional manifold (principal component) of the data by diffusion is obtained [36–38]. The specific calculation steps are as follows.

Affinity matrix or kernel: The input matrix of the DME must be a nonnegative square-symmetric affinity matrix; the Gaussian kernel is a common method used to calculate the affinity matrix. Here, we input the structural data and the FC matrices, which meets the requirements of the input matrix after processing.

Construct the diffusion matrix: (5) $$P=D^{-1}K$$(5)

Here, $K$ is obtained from normalizing the affinity matrix and alpha determines the normalization method, controlling the influence of the density of sampling points on the manifold. We set alpha at 0.5 when the diffusion is equivalent to the Fokker–Planck scattering; $D$ is the degree matrix derived from $K.$

Calculate the eigenvalues and eigenvectors of the diffusion matrix: A set of descending eigenvalues and normalized eigenvectors are obtained by spectral decomposition.

Mapping to the diffusion space: At time t, only the first d eigenvalues and eigenvectors are kept; where t controls the scale of the eigenvalues of the diffusion operator, we set t = 0 to retain the global relationship between the data points in the embedded space and d = 6 as an estimated dimension of the embedded space. The diffusion distance of ${Y^{\prime }}_i$ is computed as follows: (6) $${Y^{\prime }}_i=\left(\begin{array}{c}{\lambda }_1^t{\psi }_1\left(i\right)\\ {\lambda }_2^t{\psi }_2\left(i\right)\\ ⋮\\ {\lambda }_n^t{\psi }_n\left(i\right)\end{array}\right)$$(6)

For the adjacency matrix of $234\times 234,$ the DME yields a feature matrix of $234\times 6,$ and finally, as in SPN, we obtain the corresponding DME features of length 6.

2.4. Discrimination based on the SPN features and diffusion components

In this study, we attempted to distinguish SCH patients from healthy participants using MRI connectomes. As schematically shown in Figure 1, we extracted both the SPN and DME features of the FC and SC network. Then, we tried to put SPN and DME features, or their combination, into a SVM classifier [39–41] to train it. Finally, we obtained the test results and evaluated the performance of the different features in the discrimination. For the SPN part, the SPN filters obtained in the training process were also used to extract the corresponding SPN features of the testing dataset. Generally, we used both single mode features and fused features. We adopted two strategies when combining the features, internetwork fusion (FC and SC network) and intranetwork fusion (SPN and DME spatial features) because they may be complementary features for each other. For each type of feature, a SVM classifier was introduced for learning the distribution of the features [42,43], and a grid search approach was applied to determine an optimized set of parameters. The trained SVM classifiers were adopted to differentiate the testing dataset based on the corresponding features and to obtain the accuracies.

Figure 1. Procedures for SCH recognition based on multidimensional spatial feature fusion.

2.5. Evaluation indices

Considering that the size of the datasets is small, we adopted the leave-one-out cross-validation method to evaluate the training and testing results of the different features [44]. This means that in each training process, one subject is used as the test sample, and the remaining are used as the training samples. This process is repeated for each subject to ensure that the data of all the subjects are used alternately for testing and training. Based on the different features, the classification prediction effect under different features can be estimated. In this study, the ACC and F1-score, are used to quantify the classification performance: (7) $$\mathrm{ACC}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}\times 100\mathrm{\%}$$(7) (8) $$\mathit{precision}=\mathrm{TP}/(\mathrm{TP}+\mathrm{FP})$$(8) (9) $$\mathit{recall}=\mathrm{TP}/(\mathrm{TP}+\mathrm{FN})$$(9) (10) $$\mathit{F}1\mathit{\text{-s}}\mathrm{core}=\frac{2\mathit{\text{\hspace{0.17em}*\hspace{0.17em}precision*recall}}}{\mathit{precision}+\mathit{recall}}\times 100\mathrm{\%}$$(10) where TP is the number of true positive samples, TN is the number of true negative samples, FP is the number of false positive samples, and FN is the number of false negative samples.

3. Results

Based on the functional and SC obtained from the MRI connectomes, we characterized the spatial features of brain connectomes of the SCH patients using SPN and DME methods, and based on this, we attained the classification of patients with SCH and the healthy subjects. First, based on the SPN features and DME features of FC network and SC network, the classification performance of the SCH patients and healthy subjects under the two methods were tested. It is not difficult to observe from Table 1 that different feature numbers will influence their classification performance, and the classification performance based on the FC and SC matrices is also different. Note that the bold values present the best performance of one method.

Table 1. The performance of classifying SCH patients based on single mode features.

Feature	No. of the features	FC	SC
SPN	1 (pairs)	67.50%	60.00%
	2 (pairs)	77.50%	56.25%
	3 (pairs)	77.50%	56.25%
	4 (pairs)	77.50%	60.00%
	5 (pairs)	75.00%	61.25%
	6 (pairs)	72.50%	60.00%
	7 (pairs)	68.75%	55.25%
	8 (pairs)	63.75%	56.25%
DME	1	62.50%	63.75%
	2	62.50%	65.00%
	3	65.00%	60.00%
	4	55.00%	72.50%
	5	55.00%	68.75%
	6	56.25%	63.75%

Based on the FC network and the SC network, the differences depicted by the SPN features and the DME features are different, and there may be information complementarity. We try to fuse the FC and SC network features, and the SPN and DME features, in the same network to explore the impact of the different fusion methods on the classification accuracy. Note that we control the number of the SPN and DME feature to the same for better comparison. First, based on the internetwork fusion of the FC network and SC network, we explored the performance of their SPN features and DME features in the classification of SCH patients and healthy subjects, respectively. The results are shown in Table 2, the bold values present the best performance of one method. It is not difficult to observe that the recognition accuracy of SCH patients and healthy subjects based on the SPN method is improved to 80%, with 81.27% F1-score. However, the recognition of patients with SCH, based on fused DME features has not been effectively improved, and the highest recognition accuracy is only 65.00%.

Table 2. The performance of classifying SCH patients based on internetwork fusion features.

		No. of FC features	No. of SC features
			1	2	3	4	5	6
SPN	ACC	1	56.25%	61.25%	68.75%	76.25%	75.00%	70.00%
		2	61.25%	57.50%	66.25%	73.75%	71.25%	71.25%
		3	68.75%	66.25%	67.50%	77.50%	73.75%	71.25%
		4	76.25%	73.75%	77.50%	80.00%	78.75%	73.75%
		5	75.00%	71.25%	73.75%	78.75%	75.00%	75.00%
		6	70.00%	71.25%	71.25%	73.75%	75.00%	72.50%
	F1-score	1	48.75%	54.35%	57.46%	75.13%	66.03%	69.60%
		2	54.35%	56.56%	59.68%	76.56%	68.86%	73.15%
		3	57.46%	59.68%	58.40%	75.13%	67.63%	71.11%
		4	75.13%	76.56%	75.13%	81.27%	78.55%	77.05%
		5	66.03%	68.86%	67.63%	78.55%	70.29%	75.55%
		6	69.60%	73.15%	71.11%	77.05%	75.55%	73.99%
DME	ACC	1	65.00%	58.75%	57.50%	65.00%	61.25%	58.75%
		2	63.75%	58.75%	56.25%	60.00%	55.00%	56.25%
		3	57.50%	57.50%	55.00%	58.75%	58.75%	53.75%
		4	60.00%	57.50%	56.25%	61.25%	58.75%	57.50%
		5	57.50%	56.25%	52.50%	56.25%	55.00%	52.50%
		6	60.00%	57.50%	58.75%	61.25%	56.25%	53.75%
	F1-score	1	69.60%	65.24%	64.35%	71.74%	67.63%	64.17%
		2	66.58%	63.06%	61.46%	65.57%	60.84%	62.39%
		3	62.50%	62.50%	60.90%	63.62%	63.62%	58.89%
		4	64.29%	62.39%	61.50%	66.63%	63.63%	62.44%
		5	61.06%	61.83%	58.00%	62.19%	60.86%	58.24%
		6	63.71%	60.79%	63.17%	64.07%	60.19%	57.24%

Moreover, the feature based on intranetwork fusion was also tested, and the impact of its results on the patient recognition accuracy of SCH patients was also tested. The results are shown in Table 3, the bold values present the best performance of one method. Based on the feature fusion method in the same network, the recognition accuracy of patients with SCH is significantly improved. The fusion features based on the FC network achieved the highest recognition accuracy of 82.50%, and the recognition accuracy based on the SC network fusion feature reached 87.50%, and the corresponding F1-score of 84.11% on FC network fusion and 87.94% on SC network fusion. Note, that since the dataset is a resting-state MRI connectomes dataset, it is difficult to reflect the differences in the finely divided brain functional networks of SCH patients during task execution. Therefore, the fused features’ recognition effect based on SC is superior to that based on FC.

Table 3. The performance of classifying SCH patients based on intranetwork fusion features.

		No. of SPN features	No. of DME features
			1	2	3	4	5	6
FC	ACC	1	53.75%	53.75%	53.75%	52.50%	52.50%	53.75%
		2	57.50%	60.00%	60.00%	60.00%	58.75%	56.25%
		3	65.00%	68.75%	71.25%	71.25%	73.75%	72.50%
		4	76.25%	77.50%	81.25%	82.50%	82.50%	80.00%
		5	76.25%	72.50%	75.00%	75.00%	76.25%	76.25%
		6	70.00%	68.75%	68.75%	68.75%	67.50%	67.50%
	F1-score	1	47.08%	47.08%	47.08%	42.62%	42.62%	47.08%
		2	55.68%	60.62%	58.84%	55.59%	55.28%	56.36%
		3	59.58%	60.32%	65.33%	67.10%	67.06%	65.15%
		4	78.77%	80.23%	82.96%	83.40%	84.11%	79.94%
		5	71.25%	67.06%	72.17%	71.34%	74.01%	74.01%
		6	70.53%	70.06%	70.61%	70.61%	68.95%	68.95%
SC	ACC	1	57.50%	57.50%	52.50%	53.75%	56.25%	55.00%
		2	62.50%	61.25%	60.00%	63.75%	65.00%	65.00%
		3	72.50%	72.50%	72.50%	73.75%	73.75%	75.00%
		4	82.50%	83.75%	86.25%	87.50%	82.50%	86.25%
		5	80.00%	81.25%	78.75%	78.75%	83.75%	83.75%
		6	76.25%	75.00%	72.50%	72.50%	73.75%	71.25%
	F1-score	1	44.72%	44.72%	37.33%	42.10%	43.61%	40.63%
		2	61.39%	60.59%	60.40%	63.06%	63.06%	62.09%
		3	64.56%	64.56%	64.56%	64.89%	64.89%	66.29%
		4	82.83%	84.29%	86.52%	87.94%	84.26%	87.22%
		5	76.80%	79.49%	75.97%	75.97%	82.80%	82.26%
		6	76.71%	75.38%	73.99%	73.05%	74.95%	70.38%

Finally, we compared our proposed method with other existing methods on the same dataset (i.e. the 27_SCHZ_CTRL_dataset), and the results are summarized in Table 4. In the original study that provided this dataset, the authors performed a recursive feature elimination and support vector machines (RFE-SVM) approach to identify the most meaningful biomarkers from the structural, functional, and multi-modal connectomes of healthy controls and patients. As we see from Table 4, their RFE-SVM approach achieved the average accuracy of around 76% by fusing multimodal connectome [14], which is far less than our reported result (87.50%). In addition, we further conducted experiments on this dataset by using two common classical dimensionality reduction methods, including the principal component analysis (PCA) and isometric mapping (ISOMAP). As a linear transformation-based method, PCA tries to find the principal components of the data and projects them onto a new low-dimensional space. In contrast, ISOMAP is a manifold learning-based method that identifies the manifold structure of the data and preserves it in the new low-dimensional space. In experiments, we applied these two methods for both the FC and SC matrixes, and obtained their highest accuracies and the corresponding F1-scores. The results presented in Table 4 clearly demonstrated that our method outperforms both the PCA and ISOMAP methods in terms of the accuracy and F1-score. Together, our above findings further showed the effectiveness of our method based on multidimensional spatial feature fusion for identifying SCH patients.

Table 4. Comparison with other methods on the same dataset (27_SCHZ_CTRL_dataset).

Method	ACC	F1-score
RFE-SVM [14]	∼76%	–
PCA	71.25%	70.17%
ISOMAP	77.50%	75.45%
Our method	87.50%	87.94%

4. Discussion

In this paper, we use the SCH MRI connectomes dataset and combine two methods, CSP and diffusion mapping embedding, to explore the influence of multidimensional spatial network features on the pattern recognition of SCH. For this purpose, we separately employed the CSP and DME methods to extract the features of functional network and structural network. Moreover, based on the functional network and structural network constructed from the rs-MRI connectomes, by comparing the fusion between the networks of the functional network and the structural network, as well as the fusion of the different features under a single network, we explored the effects of the different network feature fusion methods on pattern recognition of brain diseases. In the SCH MRI connectomes dataset of the Zenodo platform, by fusing the SPN features and DME features of the SC network, the recognition accuracy of the SCH patients and healthy subjects was improved to 87.50%, which is significantly superior to the recognition accuracy of approximately 76% achieved by Gutiérrez-Gómez et al. based on the SC network full connectivity feature [14]. At the same time, we obtained a corresponding F1-score of 87.94%, indicating that the model has a good balance between precision and recall, and can accurately classify SCH patients and healthy subjects.

The SPN features or DME features based on the FC network or SC network can improve the recognition of SCH to a certain extent but the specific classification performance is different. In general, the classification performance of SPN features based on the FC network, and DME features based on the SC network, is better. Different classification performances indicate that SCH patients and healthy subjects have differences in different networks or have different characteristics. Figures 2 and 3 show the SPN filters extracted from FC or SC networks and the extracted first-dimensional DME characteristics, of which the weights are reflected in each brain region. The first pair of SPN filters calculated from the functional network indicated that the difference between the SCH patients and healthy subjects is largely reflected in the posterior cerebral region and the frontoparietal cortex. The FC of the posterior cerebral region of patients with SCH with other brain regions is weaker than that of healthy subjects, and the FC of the brain regions in the frontoparietal cortex is generally stronger than that of healthy subjects. The DME features indicated that the FC gradient of many brain regions in the SCH patients is significantly different from that in the healthy subjects, mainly reflected in the left and right prefrontal cortex, and temporal pathway. Although the two features both reflect some key brain regions, the specific features they reflected are not identical, so the two features based on the same network were complementary in their formation information.

Figure 2. SPN filter topology structure and DME first dimensional features based on functional network. (a) SPN filter; (b) DME feature.

Figure 3. SPN filter topology structure and DME first dimensional features based on structural network. (a) SPN filter; (b) DME feature.

Multidimensional feature fusion can make use of the complementary advantages of heterogeneous information to make the learning of samples more comprehensive and effective. We explore the impacts of two different feature fusion approaches, internetwork fusion based on FC, and SC features and intranetwork fusion based on the SPN and DME features in a single network, on the recognition performance. In summary, both internetwork fusion and intranetwork fusion can improve the recognition accuracy to a certain extent but the specific effect depends on the representation effect of each feature on the network, and the degree of information redundancy between the fused features. Both the SPN and DME features based on FC or SC networks have good classification performance. The intranetwork feature fusion outperforms the internetwork fusion in recognition performance, with a recognition accuracy of 87.50%. Our results further show that a reasonable feature fusion method can take advantage of the complementary advantages of multidimensional feature information to achieve a higher classification accuracy, while an unreasonable scheme may lead to an amplification of the interference and can affect the accuracy. For many brain diseases, the specific complex mechanisms are not fully understood. In the case that the characterized information cannot accurately depict the pathological differences, the fused high-dimensional features are beneficial to improving the fitting precision of the training set.

In this paper, we achieved high-precision recognition of SCH patients by multidimensional spatial feature fusion, but there are still some research shortcomings needed to be further improved. First, for the selection of datasets, this paper only selects an SCH MRI connectomes dataset and does not extend the research conclusions to more datasets. Due to the complex mechanism of brain diseases and the different changes in brain activity and structure caused by different brain diseases [45–48], there needs to be further exploration of whether the network analysis and network feature extraction in this paper are also applicable to other brain diseases. In addition, this work was performed on relatively small datasets, and the stability of the results needs to be further explored using larger sample datasets. At the same time, the data validation in this paper is mostly a single-site study, and the samples involved make it difficult to ensure the representation of the corresponding population. In general, multisite research that combines datasets from different medical or research institutions is beneficial for verifying the robustness and universality of the results, and more systematic inquiries can be considered in the future.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was supported by the STI 2030-Major Project (Grant No. 2022ZD0208500), the Sichuan Science and Technology Program (Grant No. 2023NSFSC1595), and the Hospital Fund of Sichuan Provincial People’s Hospital (Grant No. 2022QN15).

Notes on contributors

Shuqi Guo

Shuqi Guo, she is currently a PhD student of School of Life Science and Technology, University of Electronic Science and Technology of China. Her current research interests include pattern recognition, brain dynamics and multimodal brain image analysis.

Yuhang Lin

Yuhang Lin, he graduated from School of Life Science and Technology, University of Electronic Science and Technology of China in 2022. His research direction during master's degree is multimodal neural signal analysis and pattern recognition.

Shi Zhao

Shi Zhao, she graduated from School of Life Science and Technology, University of Electronic Science and Technology of China in 2023. Her research direction during master's degree is multimodal neural signal analysis.

Yan Cui

Yan Cui, he is currently with the Department of Neurosurgery, Sichuan Provincial People's Hospital. His research interests include pattern recognition, brain dynamics and multimodal neural signal analysis.

Yang Xia

Yang Xia, she is currently a full Professor with the Clinical Hospital of Chengdu Brain Science Institute, MOE Key Lab for Neuroinformation, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China. Her current research interests include visual electrophysiology and computational neuroscience.

Ke Chen

Ke Chen, he is currently an associate professor at the School of Life Science and Technology, University of Electronic Science and Technology of China. His current research interests include visual electrophysiology and computational neuroscience.

Dezhong Yao

Dezhong Yao (Senior Member, IEEE), he is currently a full Professor of neuroengineering and neurodata with the University of Electronic Science and Technology of China (UEST C), Chengdu. He is also the Dean of the Sichuan Institute for Brain Science and Brain-Inspired Intelligence, Chengdu. His research interests include EEG, simultaneous EEG and functional magnetic resonance imaging (fMRI), and brain--apparatus communication.

Daqing Guo

Daqing Guo, he is currently a full Professor with the Clinical Hospital of Chengdu Brain Science Institute, MOE Key Lab for Neuroinformation, School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China. His current research interests include computational neuroscience, brain-inspired intelligence and digital twin brains.