This research recruited 114 patients with motor dysfunction after stroke. The inclusion criteria were: (1) diagnosis by CT or MRI; (2) first onset of the disease; (3) duration of 2 weeks–3 months; (4) lesions confined to a single hemisphere, primarily involving the basal ganglia and their adjacent areas; (5) age range of 40–75 years old; (6) right-handedness; and (7) stable vital signs and clear consciousness. The exclusion criteria were: (1) history of stroke; (2) history of neurologic or psychiatric disorders, as well as other serious physical diseases; (3) contraindications to MRI examination. Based on the exclusion criteria, four patients were excluded due to previous neurological or mental disorders, and two patients with excessive head movement (translation > 2.5 mm, movement > 2.5°, or mean FD > 0.5 mm) were also excluded. Eventually, 108 patients were incorporated. According to previous research suggestions (Pendlebury et al., 2011; Sachdev et al., 2004), stroke patients were divided into two subgroups: the patients with MMSE score ≥ 26 and MoCA scores ≥ 26were classified as the PSCD group (n = 75), while those with MMSE scores < 26 and MoCA scores < 26 were classified as the PSCMD group (n = 33). Furthermore, 35 healthy individuals who were matched by age and gender were recruited in this research (Figure 1). This study was approved by the Ethics Committee of the 900 Hospital of the People’s Liberation Army Joint Logistics Force (NO. 2015011). Every participant provided written informed consent.

With the application of the Fugl-Meyer assessment scale (FMA), motor function was evaluated in this study (Chen et al., 2022). Besides, daily life abilities were evaluated using the Modified Barthel Index (MBI) (Ohura et al., 2017). The Brief Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCa) were employed in the assessment of cognitive function (Folstein et al., 1975; Nasreddine et al., 2005).

A 3.0T Siemens Trio scanner (Erlangen, Germany) with a 12-channel phased-array head coil was employed for the whole-brain imaging. Participants were positioned supine, and foam padding was applied to stabilize the head and minimize the motion artifacts. Throughout the scan, subjects were asked to keep their eyes closed and stay awake, maintain relaxed breathing, and avoid voluntary movements. rs-fMRI and detailed 3D T1-weighted structural scans were achieved from all the participants. Gradient-echo planar imaging was employed in the rs-fMRI protocol, and the parameters were set as: repetition time (TR) = 2,000 ms, echo time (TE) = 21 ms, 33 contiguous axial slices (4 mm thickness, 0.8 mm interslice gap), matrix = 64 × 64, field of view = 240 × 240 mm², and a total of 180 volumes were collected over 6 min. 3D T1-weighted parameters were parameterized as follows: TR = 1,900 ms, TE = 2.52 ms, layer thickness = 1 mm, no layer spacing, field of view = 240 × 240 mm², matrix = 256 × 256, number of layers = 176.

Lesions were meticulously delineated layer by layer on T1-weighted MR images utilizing MRIcron¹. Subsequently, the regions of interest (ROIs) were normalized to the MNI spatial template through MR fragment normalization within the SPM8 clinical toolbox. Finally, lesion maps from stroke patients were binarized and overlaid onto T1-weighted templates in the MRIcron software to generate lesion probability maps. We constructed a composite ROI by integrating the normalized ROIs and superimposed this composite ROI onto the T1-weighted template to illustrate the overlapping regions (Figure 2).

Before data preprocessing, we flipped the MRI data of patients with left-sided lesions from left to right along the midsagittal line. DPABI² was employed for preprocessing by the following steps: (1) Transform the initial DICOM file into the 4DNifti file format; (2) Remove the first 10 time points, alongside the concurrent slice-time correction and head motion correction on the remaining 170 time points, excluding participants with translation > 2.5 mm, movement > 2.5°, or mean FD > 0.5 mm; (3) Co-register the realigned image with a single 3D T1-weighted image to complete the transformation from a single space to the standard Montreal Institute of Neuroscience space; (4) Smooth on a 6 mm half-height, full-width Gaussian space; (5) Remove noise by regression including Friston’s 24 motor parameters, white matter, cerebrospinal fluid, and global brain signals; (6) Filter in the frequency range of 0.01–0.10 Hz to mitigate the interference of high-frequency noise and low-frequency drift on the images.

Hidden markov model is a markov process with the incorporation of hidden states. The changes in brain activity over time scales can be explained by a limited number of hidden states. Firstly, the AAL90 brain atlas was employed as the later observation sequence in the extraction of the time series of the subjects after preprocessing. Secondly, 90 ROIs × 170 time-point spatio-temporal data were generated for the construction of HMM. Subsequently, the HMM-MAR toolbox in MATLAB³ was used to define states through multivariate Gaussian distributions (Vidaurre et al., 2017). Iterative calculations ranging from 2 to 15 states were conducted following a prior study (Moretto et al., 2022). Finally, the number of different HMM states was evaluated by summarizing statistical indicators, which were composed of the minimum free energy and the median fraction occupancy. It was found that the HMM reached the minimum free energy state when the number of states was 5, and 5 HMM states were selected for subsequent studies (Figures 3A–C).

The metrics reflecting temporal brain dynamics were examined: (1) FO: the ratio of time that each subject spent in a specific state. (2) LT: LT is the stretch that a state lingers before shifting into another, which could reflect the stability of the state; (3) Switch rate (SR): it is defined as the frequency of transitions between different states, which could indicate the speed of dynamic changes; (4) TP is the key factor of HMM, which represents the likelihood of changing from a state to another (Figure 3D).

The SPSS 23.0 software (SPSS Inc., Chicago, Illinois, United States) was utilized in this study. Continuous variables were expressed as mean ± standard deviation (SD), while categorical variables were presented as counts. The Shapiro-Wilk test was employed in the assessment of the normality of the data. Continuous variables were assessed by independent t-tests or one-way analysis of variance (ANOVA), while categorical variables were analyzed with chi-square tests. A two-tailed t-test was applied in the comparison of FO, LT, and SR of HMM status in stroke and HCs, as well as PSMD and PSCMD patients. False discovery rate (FDR) correction was applied to multiple comparisons. p < 0.05 threshold was considered to indicate statistical significance.

A non-parametric permutation test was utilized to compare HMM state transition probabilities between stroke patients and HCs, as well as PSMD and PSCMD, with the involvement of 5,000 permutations. Moreover, Spearman correlation analysis was employed in the evaluation of the connections between HMM state-time attributes and FMA and MMSE scores. p < 0.05 threshold was considered to indicate statistical significance.

To validate the robustness of the results obtained from the HMM approach, we perform a validation analysis using the sliding window analysis method and K-means clustering framework. Firstly, the dynamic functional network connectivity (FNC) was computed using a sliding time window approach with the window size set to 22 TR (44s) and a step size of 1 TR. Secondly, a K-means clustering algorithm was applied to the windowed FNC matrix to extract the same number of states as in the HMM model (K = 5). Finally, the temporal properties of the dynamic FNC states were extracted to assess the consistency of the population trends.

A total of 108 ischemic stroke patients and 35 HCs were included in this research. Regarding the aspects of age and gender, no significant differences were observed between the stroke patients and HCs. In stroke subgroup analysis, the MoCA scores and MMSE scores were significantly different between PSMD and PSCMD (p < 0.05), whereas no differences could be found in the aspects of stroke duration, type, lesion hemisphere, FMA, and MBI (Table 1).

As shown in Figure 3C, State1 is mainly a decrease in activation for most regions, but its average functional activity is relatively higher than that of state 4. State2 is mainly a relative increase in activation for most regions, but its average functional activity is relatively lower than state 5. State 3 is mainly characterized by relatively decreased activation in most areas, but increased activation in the right frontal_mid, right frontal_mid_orb, right temporal_sup, right inferior parietal lobule, right coronal gyrus, and right anterior cuneus. State4 is primarily a decrease in overall regional activation. State5 is primarily an increase in overall regional activation.

Compared with HCs, stroke patients exhibited noticeably higher FO (t = 2.603, p = 0.018, FDR-corrected) and LT (t = 2.255, p = 0.040, FDR-corrected) in State 2, as well as FO (t = 3.025, p = 0.008, FDR-corrected) and LT (t = 2.903, p = 0.010, FDR-corrected) in State 3; whereas FO (t = −4.732, p = 0.0001, FDR-corrected) and LT (t = −4.175, p = 0.0002, FDR-corrected) in State 5 exhibited a significant decreasing trend. Furthermore, it was revealed by subgroup analyses that PSCMD patients showed significantly higher FO (t = 3.096, p = 0.006, FDR-corrected) and LT (t = 2.332, p = 0.022, FDR-corrected) in State 3 than PSMD patients (Figure 4).

According to the comparison of the transfer patterns of HMM states, no significant difference in SR could be observed (p > 0.05) (Figure 5A), which could suggest that stroke patients exhibited stable network dynamics (similar to HCs) during resting-state scanning. However, the TP of HMM states varied between the two groups (Figure 5B). Specifically, compared with HCs, stroke patients exhibited a significantly higher transition probability for state 1 to state 3, state 2 to states 1–3, state 3 to states 1–3, and state 4 to state 1 (state 1 to 3: p = 0.011; state 2 to 1: p = 0.031; state 2 to 2: p = 0.010; state 2 to 3: p = 0.001; state 3 to 1: p = 0.040; state 3 to 2: p = 0.007; state 3 to 3: p = 0.011; state 4 to 1: p = 0.025). Besides, stroke patients exhibited a significantly lower probability of transitioning from HMM states 1, 4, and 5 to state 5 (state 1 to 5: p = 0.002; state 4 to 5: p = 0.001; state 5 to 5: p = 0.00001). Overall, it could be implied by these results that stroke patients exhibited a noteworthy pattern of aberrant transitions between HMM states.

In the stroke subgroup, no discernible variation was observed in SR among the HMM statuses (Figure 5C), while significant differences could be found in TP between the PSMD and PSCMD groups. Compared with PSMD, patients with PSCMD show significantly higher transition probabilities from state 3 and state 5 to state 3 (state 3 to 3: p = 0.013; state 5 to 3: p = 0.019) (Figure 5D).

The spatial activation patterns of large-scale whole-brain network states were investigated in this study. Compared to stroke patients, HCs were mainly characterized by state 5 (Figure 6A). State 5 showed that regions with enhanced activation were mainly located in the precentral gyrus, the postcentral gyrus, the supplementary motor area, the superior frontal gyrus, the middle frontal gyrus, the precuneus, the subparietal lobule, the subcortical regions (thalamus, caudate nucleus, and nucleus of the bean-shell), and the hippocampus. Conversely, areas with weakened brain activation were predominantly in the cuneate lobes.

According to the results, stroke patients were predominantly in states 2 and 3 (Figures 6B, C). Specifically, the brain regions with enhanced activation in state 2 were mainly composed of the precentral gyrus, supplementary motor area, superior frontal gyrus, middle frontal gyrus, precuneus, subcortical regions, middle occipital gyrus and inferior occipital gyrus, and cuneus, while the activation levels were significantly lower than state 5. The regions with weakened activation in state 3 were mainly in the middle frontal gyrus, superior frontal gyrus, precentral gyrus, left supplementary motor area, subcortical regions, middle temporal gyrus, inferior temporal gyrus, angular gyrus, middle occipital gyrus and inferior occipital gyrus, and insula, whereas those with enhanced activation were predominantly in the right superior frontal gyrus, right middle frontal gyrus, right inferior parietal lobule, right angular gyrus, and right precuneus. Regarding the stroke subgroup analysis, PSMD patients were mainly dominated by state 2, while PSCMD patients were mainly dominated by state 3.

A negative correlation could be found between FO and severity of cognitive impairment in state 3 of PSCMD patients (r = −0.389, p = 0.025) (Figure 6D), while the temporal dynamic characteristics of other states exhibited no significant connections with motor or cognitive functions.

Sliding window analysis and K-mean clustering were used to obtain five states, and between-group differences in the temporal attributes of these states were calculated. We found that stroke patients had higher FO in state 3 (t = 2.933, P = 0.010, FDR-corrected), but shorter FO in state 5 (t = −4.732, P = 0.0001, FDR-corrected) compared with HCs. The comparison between stroke subgroups showed that the FO of PSCMI patients in state 3 was higher than that of PSMD patients (t = 2.436, P = 0.017, FDR-corrected). In addition, there was no difference in number of transitions between stroke and HCs group as well as the stroke subgroup in the analysis (P > 0.05). These are similar to the HMM results. However, we did not find a difference between stroke and HCs in state 2 in the validation analysis (Supplementary Figure 1).