We screened patients from the First Affiliated Hospital of Beijing University of Chinese Medicine. Criteria for inclusion comprised the following: (1) age 18–80 years; (2) patients were diagnosed with ischemic stroke and suffered for the first time (Referring to the Chinese Guidelines for the Diagnosis and Treatment of Acute Ischemic Stroke 2018 https://www.medsci.cn/guideline/show_article.do?id=fec5a1c0016ea8f7); (3) the illness duration ranged within 1.5 months; (4) motor dysfunction after stroke was the prominent manifestation; (5) patients were right-handed before the stroke. We recruited 35 demographically matched healthy controls (HCs) from the local community with no psychological or neurological disorders. The studies involving human participants were reviewed and approved by Medical Ethics Committee, Dongzhimen Hospital, Beijing University of Chinese Medicine (No: 2022DZMEC-111-02). Written informed consent was provided for the patients or their families/legal guardians.

To reduce the impact of potential confounding factors, we adopted a set of exclusion criteria: (1) inability to complete clinical scales; (2) any history of neurological disorders, such as multiple infarcts, hemorrhage, or psychiatric disorders; (3) any contraindications to undergo Magnetic Resonance Imaging (MRI).

All resting-state functional imaging scans were acquired with a 3.0-Tesla scanner (Siemens, Verio, Germany) at the MRI Unit of the First Affiliated Hospital of Beijing University of Traditional Chinese Medicine, using a single-shot T2*-weighted gradient echo planar imaging (EPI) sequence, with the following parameters: Repetition Time (TR) = 2000 ms, Echo Time (TE) = 30 ms, fip angle = 90°, slice thickness = 3.5 mm, skip = 1 mm, and slice number = 31, matrix size = 64 $\times$ 64, field of view = 225 $\times$ 225 mm².

The MRI images of 26 patients with right hemispheric lesions were flipped along the midsagittal plane before data analysis [27]. Data preprocessing was implemented using SPM12 software (Statistical Parametric Mapping 12, http://www.fil.ion.ucl.ac.uk/spm). Each functional Magnetic Resonance Imaging (fMRI) scan consisted of 179 time points. We discarded the first ten volumes to correct magnetic field equilibration, signal stabilization, and adaptation of subjects to scanner noise. Subsequent data preprocessing steps include slice time correction, motion correction, spatial normalization, and smoothing. Next, we standardized all images to the Montreal Neurological Institute (MNI) space. The resampling voxel size was 3 $\times$ 3 $\times$ 3 mm. Finally, we applied spatial smoothing with a 6 $\times$ 6 $\times$ 6 mm Gaussian kernel and bandpass temporal filtering (0.01–0.1 Hz) to eliminate high-frequency noise in the normalized images.

The motor function was assessed using the Fugl-Meyer Assessment (FMA). FMA serves as a robust tool for evaluating disease severity, rehabilitation progress, and treatment response [28], which can be subdivided into FMA upper extremity scores (FMA-U) and FMA lower extremity scores (FMA-L). A lower FMA score indicates more severe motor impairment caused by stroke. Considering the diverse range of motor impairment severity among the enrolled patients, additional subgroup analyses were conducted based on the FMA score. According to Dawson et al. [29], patients with an FMA score $\ge$50 were designated mild-to-moderate motor impairment, while those with an FMA score 50 were classified as severe motor impairment. Patients underwent motor function assessment twice: the first time was on the day of enrollment, and the second time was two weeks later. During this period, patients received standard treatment in the Department of Neurology.

In this investigation, we restricted the stroke group to patients with strokes lasting less than 1.5 months, with unilateral limb dysfunction as the primary symptom. Additionally, we enrolled patients with ischemic lesions localized on one side of the centrum semiovale, corona radiata, and basal ganglia regions to ensure a high degree of sample homogeneity. In addition, we explained the scanning procedure to the participants before the MRI scan. We also emphasized the importance of maintaining audiovisual closure and avoiding cognitive processes during scanning. Participants displaying excessive head motion ($>$3 mm of displacement or $>$3° of rotation in any direction) were systematically excluded from the study.

We utilized the DynamicBC toolbox (http://www.restfmri.net/forum/DynamicBC) [30] to generate dynamic FC. The entire brain was divided into 90 regions (excluding the cerebellum) using the Anatomical Automatic Labeling (AAL) atlas [31]. This atlas divides the brain into five Resting-State Networks (RSNs), including the Default Mode Network (DMN), Attention Network (AN), Visual Network (VN), Subcortical Network (SN), and Sensorimotor Network (SMN). Dynamic FC was computed through the sliding time window method, which involved segmenting the entire fMRI time series into smaller time windows and calculating the connectivity matrix for each time window. Then, dynamic multilayer brain networks were constructed by linking a node in the coherence matrix to its corresponding node in the coherence matrix of adjacent time windows. Different from methods that treat the data in each window as an independent sample, the multilayer network approach considers the temporal dependence between windows. Following the methodology of previous studies [14, 32], we employed a sliding window length of 50, with the window shifted by 1 step. FC between brain regions was assessed using Pearson correlation coefficients and Fisher Z-transformed. Consequently, for each participant, a matrix of dimensions 90 $\times$ 90 $\times$ 120 (nodes $\times$ nodes $\times$ layers) was obtained, arranged in the order of scanning time.

Multilayer modular detection describes clusters of highly connected regions as modular structures that change over time. We applied the modularity maximization algorithm to detect the temporal modular within the dynamic multilayer network [33]. This algorithm identifies modular temporally by maximizing a Louvain-like modularity function ($Q$):

$$Q=\frac{1}{2\mu }\sum _{ijlr}\left(\left(A_{ijl}-{\gamma }_l\cdot P_{ijl}\right){\delta }_{l\mathrm{r}}+{\delta }_{ij}{\omega }_{jlr}\right)\delta (g_{il},g_{jr})$$

Where µ is the total edge weight in the network, $A_{ijl}$is the adjacency matrix between node $i$and node $j$ in layer $l$. The ${\gamma }_l$ is structural resolution parameter of layer $l$. The ${\omega }_{jlr}$ is the temporal coupling parameter for node $j$ between layer $l$and layer $r$. We set the two parameters to 1 to prevent bias towards a specific number of modulars or a particular scale of temporal dynamics in the modular structure [34]. $P_{ijl}$ denotes the matrix of expected weights under the Newman-Girvan null model. The modular assignments of node $i$ in layer $l$ and node $j$ in layer $r$ are represented by $g_{il}$ and $g_{jr}$. The modularity of the multilayer network is denoted by $Q$. A high $Q$ value indicates that the multilayer network is reasonably subdivided into modules. We calculated the number of modules ($N$) generated throughout the scanning process. We optimized the quality function of multilayer modularity 50 times for each temporal network since the modular detection algorithm is non-deterministic and tends to near degeneracies [35].

Multilayer modularity analysis classifies brain regions into modules for each time window. To capture temporal changes in module configurations, we defined the flexibility of node $i$ ($F_i$) as the number of times node $i$ changed its module assignment within consecutive time windows, normalized by the total number of possible changes [20]. Flexibility can be described using the following equation:

$$F_i=1-\frac{1}{T-1}\sum _{s=1}^{T-1}\delta (g_{i,s},g_{i,s+1})$$

Here, $T$ represents the time window. If node $i$ remains in the same module in both layer $s$ and layer $s+1$, then $\delta (g_{i,s},g_{i,s+1})=0$, else, $\delta (g_{i,s},g_{i,s+1})=1$.

The Module Allegiance Matrix (MAM) detection was introduced to analyze interactions between different brain areas in the multilayer networks. The values in the MAM represent the proportion of times during the scanning process that the nodes within the RSNs are assigned to the same module. If two nodes within the RSN are consistently assigned to the same module, their module allegiance is 1. If they never appear together in the same module, their module allegiance is 0. MAM is calculated as follows:

$$P_{ij}=\frac{1}{OT}\sum _{o=1}^o\sum _{t=1}^T{a}_{i,j}^{k,o}$$

Where $O$is the number of repetitions, if node $i$ and node $j$ are in the same module, then$a_{i,j}^{k,o}=\mathrm{\hspace{0.25em}1}$, otherwise, $a_{i,j}^{k,o}=0$. Module loyalty values for each RSN are obtained by averaging the module allegiance values of the nodes within each RSN, represented by an $N*N$ matrix, where $N$ is the number of brain networks.

The dynamic interaction among brain regions can be more precisely characterized using two metrics: the recruitment coefficient and the integration coefficient, derived from the MAM. The recruitment coefficient describes the probability of nodes from the same RSN reorganizing into modules during scanning, thus reflecting internal alignment. Conversely, the integration coefficient denotes the probability of a node sharing the same module with nodes from different RSNs, representing the alliance with regions from diverse RSNs.

The recruitment coefficient for node $i$ in the functional network $S$ can be calculated as follows:

$$R_i^S=\frac{1}{n_S}\sum _{j\in S}{P}_{ij}$$

The integration coefficient for node $i$ in functional network $S$ can be defined as:

$$I_i^S=\frac{1}{N-n_S}\sum _{j\notin S}{P}_{ij}$$

In the two appealing equations, $nS$ represents the number of nodes within the network $S$. It is noteworthy that the dynamic recruitment and integration of networks are not mathematically equivalent to the strength of the functional or structural connectivity of regions. Instead, they quantify the time-dependent probability that the brain region is connected to its network (recruitment) or another network (integration) over time. The flow chart of data processing in this study is shown in Fig. 1.

Fig. 1.

Data analysis flowchart. (a) Image pre-processing and time series extraction from fMRI signals by AAL atlas. (b) Sliding window analysis to construct individual dynamic functional matrix. (c) Multilayer network construction and module division. (d) Calculation of flexibility and construction of the Module allegiance matrix. (e) Calculation of recruitment and integration coefficients. Abbreviations: AAL, Anatomical Automatic Labeling; RSN, Resting State Networks; fMRI, functional Magnetic Resonance Imaging; BOLD, blood oxygen level-dependent; SMN, Sensorimotor Network; VN, Visual Network; AN; Attention Network; DMN, Default Mode Network; SN, Subcortical Network.

In the analysis of the modular structure of multilayer networks, the parameter $\gamma$can impact the size and number of modules within each layer. The parameter $\omega$ may influence the number of modules across layers. To validate our findings, we employed various analysis strategies, incorporating variations of parameter γ (0.95, 1, 1.05) and ω (0.95, 1, 1.05). Similarly, the calculations were repeated 50 times for each parameter combination. Pearson correlation analyses were performed on the results obtained from these analytical strategies to confirm the reliability of the main findings (refer to Supplementary Methods for details).

We performed Pearson correlation analyses of FMA scores on the day of the MRI scan and dynamic indicators with characteristic changes after stroke, p 0.05 indicates that the relationship between dynamic indicators and motor impairment is established.

We employed a single linear regression model to identify neuroimaging markers sensitive to the favorable prognosis of motor function after stroke. Dynamic measures exhibiting intergroup differences were used as dependent variables, while scale scores reflecting motor function improvement served as independent variables. The coefficients in the model were estimated using the least squares method. Subsequently, the regression equation was tested using an F-test. A significance level of p 0.05 was considered indicative of a sensitive predictor for motor improvement.

Differences in demographic information among the three groups were analyzed using a one-way analysis of variance (ANOVA). For clinical scales, a paired-sample t-test was employed to identify differences between the pre-assessments and post-assessments in the stroke group. Metrics related to multilayer brain networks, including Q, flexibility, recruitment, and integration coefficients, were evaluated by calculating the average value across all nodes within the same RSN. The Kolmogorov-Smirnov test was subsequently applied to assess data normality. If the data were not normally distributed, the Kruskal-Wallis H-rank sum test was used. If the data were normally distributed, ANCOVA was used, with age and gender included as covariates. A post-hoc analysis Multiple Comparison of Means - Tukey Honestly Significant Difference (Means - Tukey HSD) was conducted for dynamic multilayer network parameters showing group effects. Differences were considered significant at p 0.05 after multiple comparisons. Parametric variables are reported as mean $\pm$ standard error of the mean (SEM), whereas non-parametric variables are presented as the median with interquartile range. All statistical analyses were performed using Python 3.11.3 (https://www.python.org/downloads/release/python-3113/).

Participants included 50 stroke patients and 35 HCs. Excessive head movement was observed in 4 stroke patients and 4 HCs. Ultimately, 46 stroke patients and 31 HCs were included in the analyses. The stroke group was further divided into mild-moderate impairment (n = 30) and severe impairment (n = 16) groups. There were no significant differences in age (p = 0.333) and gender (p = 0.307) were found among the three groups. The clinical characteristics of participants are summarized in Supplementary Table 1.

Stroke patients showed significant improvement in FMA, FMA-L, and FMA-U scores, indicating improvement in motor performance. Details are provided in the Table 1.

Ischaemic lesions in stroke patients included in the study were predominantly distributed in the basal ganglia, internal capsule, thalamus, and corona radiata. Employing MRIcron software (Version 4, http://www.softpedia.com/get/Science-CAD/MRIcron.shtml), we labeled the infarct lesions of each participant. The distribution of infarct lesions was overlapped for presentation (Fig. 2).

Fig. 2.

Overlapping lesion images. The images with left hemisphere lesions were flipped from left to right along the median sagittal line. The blue lines represent the slices of the brain shown. Abbreviations: L, Left; R, Right.

To quantify the compartmentalization of potential functional modules in the brain, we calculated the modularity Q. The process of community switching over the scanning time is illustrated in Fig. 3A–C, with one subject randomly selected from each group as an example. There were no statistically significant differences in Q and $N$ among HCs, patients with mild to moderate impairment and patients with severe impairment (Fig. 3D,E).

Fig. 3.

Modular partitioning of multilayer brain networks. (A) Modular segmentation of HCs during the scanning time. (B,C) Dynamic module partitioning for mild-to-moderate and severe motor impairments, respectively. (D) Comparisons among stroke patients with mild-to-moderate and severe impaired motor function and HCs in modularity Q. (E) Difference in the number of modules among the three groups. Abbreviations: Q, Modularity Q; HCs, Health controls; MMP, Mild-to-moderate patients; SP, Severe patients.

We computed the flexibility of RSNs to characterize the dynamic reconstruction of modules. Patients with stroke exhibited higher flexibility in the DMN than HCs (p = 0.009). Post-hoc analyses revealed that patients with mild to moderate impairment showed significantly greater flexibility in the DMN than HCs (p = 0.022). No statistical differences were observed in the other networks. Further details can be found in Fig. 4D.

Fig. 4.

Module reconfiguration pattern differences. (A) MAM of HCs. (B) MAM of MMP. (C) MAM of SP. (D) The results showed that flexibility in the DMN were higher in MMP than in HCs, implying more frequent modular reconfiguration in MMP. (E,F) The results showed lower recruitment coefficients of SMN, AN, and DMN in MMP compared to HCs. Furthermore, the recruitment coefficient of AN was reduced in SP compared to HCs. These findings suggest a significant alteration in the pattern of modular reconfiguration and augmented interaction between brain networks after stroke. Single (*) and double (**) asterisks represent p 0.01 and p 0.05 after correcting for multiple comparisons, respectively. (G) The spatial distribution alterations of SMN (Blue), AN (Yellow), and DMN (Red) from different angles of the brain. Abbreviations: MAM, Module Allegiance Matrix; SMN, Sensorimotor Network; DMN, Default Mode Network; AN, Attention Network; VN, Visual Network; SN, Subcortical Network; HCs, Health controls; MMP, Mild-to-moderate patients; SP, Severe patients.

The MAM, also known as the temporal co-occurrence matrix [36], describes the relationship between RSNs. To explore patterns of networks in the dynamic reconfiguration of modules, we generated module loyalty matrices of RSNs for three groups. Each subject possessed a module loyalty matrix, and the group level was shown by averaging all subjects within the group (Fig. 4A–C). The darker color in the heatmap signifies a lower probability of network alignment, while a lighter color indicates a higher likelihood of alignment. Within the MAM, the intra-network color is brighter than the inter-network color in both HCs and patients (p 0.05, detailed data are provided in the supplementary material), suggesting that the probability of alignment among brain regions within the same RSN exceeds the probability of alignment between brain regions belonging to different RSNs during the continuous reorganization of modules. As displayed in the heat map, the intra-network color of DMN (Mild-to-moderate patients (MMP): p = 0.040, Severe patients (SP): p = 0.040), AN (MMP: p = 0.041, SP: p = 0.002), SMN (MMP: p = 0.033, SP: p = 0.026) in the stroke group became darker compared to the HC group, indicating a decrease in module loyalty of stroke patients within RSNs. Additionally, we observed increasing inter-network module loyalty in stroke patients compared to the HCs, indicating increased interaction between these networks (SP: SMN-DMN: p = 0.027; MMP: SMN-SN: p = 0.032).

Recruitment and integration coefficients characterize the alignment preferences of brain regions during the dynamic reconfiguration process by assessing the relationship between empirical functional networks (i.e., RSNs defined by a brain atlas) and actual delineated modules (i.e., the functional compartments of observed subjects computed by a modular maximization algorithm). The MAM provides a holistic view of the dynamics of brain networks. To further investigate the dynamic interactions within and between RSNs after stroke, we compared the recruitment and integration coefficients of the RSNs among the three groups. The results revealed that the recruitment coefficients of the patients, compared to the HCs, were significantly lower in the AN, SMN, and DMN (Kruskal-Wallis H, AN: p 0.001; SMN: p = 0.017; DMN: p = 0.017). Post-hoc multiple comparison analyses between groups indicated that the recruitment coefficients for patients with mild to moderate injuries were significantly lower in the SMN (p = 0.001), DMN (p = 0.028), and AN (p = 0.009). Furthermore, the recruitment coefficients for patients with severe injuries were significantly lower in the AN (p = 0.005). No statistically significant differences were observed in other RSNs. Additionally, there were no significant differences in the integration coefficients among the three groups. More details are presented in Fig. 4E,F. The distribution of networks with significant differences is shown in Fig. 4G.

We correlated the dynamic indicators of significant differences with the patients’ motor performance. We found that higher flexibility of DMN after stroke positively correlates with lower limb functional performance in patients with mild-to-moderate motor impairment (r = 0.369, p = 0.045, Fig. 5A). However, this correlation was not found in patients with severe motor impairment (r = 0.057, p = 0.833, Fig. 5D). In addition, dynamic indicators with statistical differences were entered into a linear regression model to establish the relationship between motor function recovery after two weeks and temporal features. We found that hyper-flexibility of DMN (R² = 0.200, F (1,28) = 5.798, p = 0.013, Fig. 5B) can predict improved lower limb motor function in patients with mild-to-moderate motor impairment. Additionally, higher recruitment coefficients of the AN predict worse recovery of lower extremity function in these patients (R² = 0.132, F (1,28) = 4.243, p = 0.049, Fig. 5C), indicating that lower recruitment coefficients correspond to improved lower limb motor recovery. However, other dynamic indicators in patients with mild-to-moderate impairment showed no relationship with the prognosis of motor function. Furthermore, we did not observe an association between dynamic metrics and lower limb motor outcomes in patients with severe motor deficits after stroke. The results are shown in Fig. 5E,F. Patients without lower limb motor impairment (FMA-L = 34 points) were excluded from regression analyses to avoid the ceiling effect [29]. Significant predictive relationships remained in patients with mild-to-moderate motor impairment, suggesting that the main results were stable and not significantly influenced by the ceiling effect. Detailed analysis procedures are provided in the Supplementary Results.

Fig. 5.

Regression analysis of dynamic indicators and lower limb motor function. (A) Higher flexibility of DMN positively correlates with lower limb function in patients with mild-to-moderate motor impairment. (B) Hyper-flexibility of DMN predicts improvement of lower limb outcomes in patients with mild-to-moderate motor deficits. (C) Higher recruitment coefficients of AN represent worse lower limb prognosis in patients with mild-to-moderate motor deficits. (D–F) No association was found between dynamic indicators and lower limb motor outcomes in patients with severe motor impairment. Abbreviations: AN, Attention Network; DMN, Default Mode Network; FMA-L, Fugl-Meyer Assessment-Lower Extremity.

Pearson correlations were calculated across different analysis strategies to assess the consistency of the main findings. The results indicated a high level of consistency, with correlations between different pairs of $\gamma$and $\omega$parameters. Detailed results of the validation analyses can be found in the Supplementary Results.

Flexibility describes the frequency of module rearrangements. Our results confirm that, even in the resting state, the brain module structures of stroke patients and HCs continuously transformed. We identified hyper-flexibility of DMN after stroke, suggesting stroke caused an unstable network structure and active reconfiguration within the DMN, especially in patients with mild to moderate motor impairment. Previous research has reported increased brain network flexibility during cognitive tasks, suggesting it responds to behavioral changes [20]. Acquiring motor skills in HCs is also closely associated with increased flexibility [37]. Disease states lead to over-flexibility, as observed in patients with autism spectrum disorder [25] and schizophrenia [38]. This remote activation may be a form of dysfunction or a compensatory response [39, 40]. The DMN is responsible for primary perception, advanced cognition, and their integration. Abnormal FC changes within and between the DMN have been reported in subcortical stroke patients [41]. Distinctive alterations in the DMN have been observed in both acute and chronic stroke patients without cognitive deficits [42]. The dynamic switching of modules in the DMN may represent an optimization strategy for delivering brain information to support diverse functional activity.

Stroke induces complex information exchange across multiple networks [43, 44, 45, 46]. Frequency-based exploration of dynamic reconfiguration in the multilayer network after stroke reveals an overly flexible state in the DMN. Probabilistic-based exploration methods, such as MAM, recruitment, and integration coefficients, complement the forms of modular reconfiguration. We observed reduced recruitment coefficients in the SMN, AN, and DMN, suggesting that stroke disrupts the original functional networks in the normal brain. Consequently, these networks are less separated from other RSNs and tend to continuously reconfigure modules with other RSNs to support neurological functions of the brain after stroke. It may serve as an adaptive response to impaired brain function. Numerous studies have consistently demonstrated altered FC within sensorimotor and cognitive control regions. The SMN has been reported to undergo significant changes in FC following acute stroke, playing a crucial role in the intricate process of motor function remodeling [47, 48]. The SMN is vital in perceptual transmission and motor initiation after stroke [49]. Anomalies in the SMN appear to mirror irregularities in motor network interaction after stroke and adaptive changes in response to motor network impairment. The AN correlates with processing incidental stimuli and attentional shifts [50]. The ventral AN assumes a critical role in accident perception and the processing of adaptive percepts [51]. The dorsal AN has been emphasized for its significant role in attentional control, visuospatial processing, and goal-directed orientation [52]. Additionally, previous research has reported FC disruptions within the AN and SMN. These abnormal changes may have implications for prognosis and recovery from stroke [53]. In summary, from the perspective of dynamic module remodeling, we observed more significant temporal variability after stroke, providing new evidence for changes in these networks.

Functional brain networks maintain communication in the resting state, exhibiting integration, interaction, and synchronization effects. Within the modular structure of the multilayer network, RSNs are assigned to different or the same modules in each layer. Specific RSNs may display abnormal dynamic modular allegiance, impairing the necessary integration of brain functions. The MAM summarizes the consistency of RSNs assigned to modules. We calculated the MAM in both stroke patients and HCs, and the results suggest that the RSNs tend to form modules with themselves in both physiological and pathological states. Notably, we observed a reduction in the MAM within all RSNs after stroke, which was evident in patients with mild to moderate and severe motor impairment. These results suggest that nodes within the same RSNs are less likely to reorganize into the same module, indicating weaker intra-network interactions after stroke. In addition, we observed an increase in the MAM between RSNs after stroke, supporting the increased interaction of networks during dynamic module reconfiguration. This increased communication between networks after stroke may reflect compensatory or maladaptive responses to impaired connectivity. The MAM can be considered a soft partitioning method to describe the degree or probability of each RSN’s participation in modules [54]. Higher module loyalty may correspond to more robust structural or functional connectivity of networks. It should be noted that dynamic network integration between networks is not mathematically equivalent to structural or functional connectivity. A network can exhibit strong structural or functional connectivity with other networks but weak dynamic network integration, suggesting its regions are less likely to combine with other networks into modules [55]. The over- and under-synchronization pattern after stroke suggests a complex, time-varying evolution of module structure. The mechanisms underlying the increase in dynamic reconfiguration after stroke remain elusive but may involve an altered excitation/inhibition balance and regulatory functions between networks.

Numerous attempts have been made to identify neural biomarkers; abnormal functional and structural connectivity were observed in patients with post-stroke motor dysfunction, which is associated with clinical impairment and recovery [10, 56]. The modular multilayer network analysis approach used in this study can sensitively capture the dynamic characteristics of the brain, compensating for the limitations of previous studies. We found over-flexibility of the DMN in patients with mild to moderate motor impairment, which reflects good motor function performance. Furthermore, multilayer network metrics with significant differences were included in the regression model to assess the relationship between these metrics and motor recovery after 2-week follow-up. The results showed that increased flexibility in the DMN may predict motor improvement of the lower limb in patients with mild-to-moderate motor function deficits. This hyperactivation might be instrumental in exchanging information during motor function recovery. RSNs with lower recruitment coefficients exhibit less uniform assignment to modules over time, while RSNs with higher recruitment coefficients indicate more dynamic interaction with other RSNs. Lower recruitment coefficients in the AN suggest that the AN is more inclined to participate in multiple brain functions and is involved in a dynamic process of reorganizing the module as a connection unit, it may also be a potential biomarker for predicting the recovery of lower extremity motor function in patients with mild-to-moderate motor deficits. The current findings suggest that early dynamic modular reorganization may reflect motor function impairment and predict recovery longitudinally. Understanding this relationship will facilitate the development of targeted and individualized treatments.

There are some limitations to acknowledge. First, although we performed a repeatability analysis to demonstrate the robustness of the results, the present study included only stroke patients with limited lesions, and the primary manifestation was unilateral motor dysfunction. Therefore, caution is needed when extending these findings to patients with infarcts in different brain regions or to patients with other post-stroke symptoms. Additionally, the relationship was not found in patients with severe motor function impairments, possibly due to slightly impaired patients benefiting more readily from re-adaptation after stroke. We would need a larger sample size of stroke patients, especially severely impaired patients, to improve the statistical power to assess the relationship between dynamic metrics and clinical scale scores.