Functional connectivity and graph theory of impaired central visual pathways in acute ischemic stroke based on functional magnetic resonance imaging

Introduction

Stroke is one of the most prevalent diseases worldwide and a significant contributor to disabilities. The primary symptoms are abnormal verbal function, impaired consciousness, and somatic movement disorders, which disrupt daily life and impose a significant economic burden on society and families (1,2). Ischaemic stroke (IS) occurs when cerebral blood circulation disorders block or significantly constrict blood vessels in a specific brain region, decreasing cerebral blood perfusion and causing ischemia, hypoxia, and brain tissue necrosis (3-5). Intravenous thrombolytics may be used to treat patients with acute ischemic stroke (AIS) within 4.5 h of stroke onset, although they can cause cerebral hemorrhage (6,7). Therefore, the development of novel therapeutic methods is an emerging area of research.

Recently, functional connectivity (FC) during the resting state has become popular following AIS and has been thoroughly examined using magnetic resonance imaging (MRI) techniques. FC reflects the interconnections between different brain regions of interest (ROIs). Most brain FC studies use static functional connectivity (sFC), dynamic functional connectivity (dFC), and group analysis of [functional MRI (fMRI)] data to generate subject-specific functional networks (8-11). AIS has been extensively studied; however, its brain mechanisms remain unknown (12).

FC can reflect the normal or abnormal function status and characterize neuronal activity between brain areas in a network (13,14). In patients with stroke, most resting-state FC studies rely on a priori selection of seed regions (15), which fails to elucidate the comprehensive characteristics of FC across the brain. Thus, brain-wide sFC research is required to understand the neurobiological mechanisms underlying these disorders. Healthy brain function requires ongoing intrinsic activity in a highly integrated network (16,17). FC has been shown to represent neuroanatomical characteristics that enable evaluation of the network’s intrinsic neural information transmission (18,19), and that resting-state brain activity spatially classifies neural activity in a specific coherent pattern. sFC, a model and data-driven algorithm based on resting-state fMRI, does not use hypothetical analyses of a priori knowledge, and whole-brain-wide ROIs of intergroup variability analysis can explain ROI interaction mechanisms. For each connection, sFC is the Pearson correlation coefficient between the time courses of the two signal components, which are produced from the ROI-based time series (20,21).

Recent studies have shown that resting-state FC is dynamic and fluctuates spontaneously over short temporal intervals. FC alterations over time can provide significant insights into cerebral function (22-24). A single resting-state fMRI session shows rapid changes in network strength and direction over time. The dynamic nature of the FC is overlooked when it is represented as a static correlation over the entire testing period. The dFC estimates FC temporal variations, preserving brain network dynamics that may affect phenotypic variance (25). Recent studies have highlighted the utility of network dynamics in predicting trait behaviors, with dFC status reflecting individual differences in clinical symptoms (26).

A scientific study indicated that dFC is a more reliable predictor than sFC (27). Additionally, dFC shows constant within-subject behavioural changes, reflecting attentional and cognitive alterations. Recent developments in “dynamic” analytical approaches have challenged the long-held concept of “static” connectivity in fMRI. These innovative methods enable researchers to observe temporal variations in cross-regional brain connections over short intervals. Consequently, connectivity changes can be evaluated within a time frame of seconds rather than that of minutes. The dFC represents brain’ states of connectivity’ and their transition pathways at a large level. These dynamic assessments can clarify spontaneous brain signal fluctuations better than static signals, which may be relevant to behaviour. dFC analysis is fundamental to the IS. With moderate motor strokes, communication across movement areas temporarily increases; however, with severe strokes, information processing is more restricted (28).

The influence of machine learning (ML) on AIS has grown substantially, driven by rapid advancements in ML technologies (29). In the field of neuroimaging, the exploration of FC has been revolutionized by ML techniques, which have provided strong methods for the comprehensive analysis of high-dimensional datasets. By employing advanced ML algorithms such as linear support vector machine (linear-SVM), radial basis function support vector machine (RBF-SVM), K-nearest neighbors (KNN), random forest (RF), and decision tree (TREE), researchers can identify complex patterns and classify functional disruptions associated with strokes, which traditional statistical methods may fail to detect (30). In response to this requirement, there has been a significant increase in the use of ML methods, motivated by a compelling reason: even a brief reduction in treatment delays can greatly improve patient outcomes.

The present study used resting-state static and dynamic FC to compare the FC between healthy controls (HCs) and patients with AIS, ensuring adequate representation from normal cognition to dementia. We used ML classification models to determine the most discriminative connection features. This study suggests that FC alterations in patients with AIS may explain cognitive decline and establish AIS-specific neuroimaging markers. We present this article in accordance with the STROBE reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-834/rc).

Methods

Participants

This study was registered with the Chinese Clinical Trial Registry (ChiCTR2100046692). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of the Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (approval number: 2021-121). All the study participants or their designated representatives provided written informed consent. Notably, individuals experiencing cognitive decline may lack the capacity to autonomously consent to participating in research studies. On their behalf, a legal guardian or representative gave informed consent for participation.

The clinical trial was conducted at the Stroke Center of Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University from March 2021 to September 2022. Patients who enrolled in the study underwent a 1-year examination and follow-up at 4 designated time points. In total, 22 enrolled cases were collected, 6 cases were discarded, and 26 healthy individuals (10 males/16 females, average age: 66.62±7.53 years) were recruited.

At present, 16 patients were followed up, and 12 who completed the necessary examinations and follow-up work were included in the study. Twelve patients (8 men/4 women, age: 66.42±10.39 years) finished the 7-day experiment and 12 patients (8 men/4 women, age: 66.42±10.39 years) completed the 3-month experiment. Among them, 12 patients who completed the 7-day experiment possessed information identical to those who completed the 3-month experiment. The follow-up time point was 3 months post-treatment.

Patient inclusion criteria: (I) age 18 to 80 years old; (II) patients with acute cerebral infarction causing damage to the central visual pathway (occipital lobe, optic cortex infarction, etc.); (III) within 7–10 days of onset; (IV) signed informed consent form. Those who are incapacitated or have limited capacity require a legal guardian to sign the informed consent form and the date of signing. Exclusion criteria: (I) hemorrhagic stroke; (II) Glasgow Coma Scale (GCS) score ≤5 (31); (III) persons with pre-existing disability [Modified Rankin Score (mRS) score >2]; (IV) women during pregnancy; (V) any condition that precludes MRI (e.g., metal object placement, claustrophobia, difficulty cooperating with unconsciousness, etc.); (VI) combination of severe functional impairment of vital organs, such as the heart, liver, lungs, and kidneys, or severe mental disorders [Hematopoietic Cell Transplant Comorbidity Index (HCT-CI) score ≥3]; (VII) combination of other serious diseases with survival <90 days; (VIII) patients with visual acuity test results suggesting previous optic nerve atrophy after enrollment (if there is unilateral eye involvement, the affected eye will be excluded); (IX) poor compliance and inability to complete the established follow-up program. Withdrawal criteria: (I) subjects who have requested to withdraw their informed consent; (II) from a medical perspective, the investigator believes it is essential for the participant to quit the study.

Neuropsychiatric assessment

Following MRI scanning, the International Cooperative Ataxia Rating Scale (ICARS) (32) of the World Federation of Neurological Disorders was used to assess the severity of ataxia in patients, with a maximum score of 100, where 0 signified normal ataxic function, and an elevated score indicated greater severity of ataxia. Furthermore, the evaluation employed the National Institutes of Health Stroke Scale (NIHSS) (33,34) and the Montreal Cognitive Assessment (MoCA) (35).

Data acquisition

A SIEMENS MAGNETOM Prisma 3.0T scanner was used for the scanning. Before scanning, the patients were checked for metal objects using a metal detector and admitted to the MRI scanning equipment room if they had no metal objects. The participants were directed to lie supine in a stable position on the MRI scanning bed and use a soft cushion to secure the sides of the head, thereby reducing head movement. The subjects were advised to maintain their composure, close their eyes, and avoid falling unconscious during the image acquisition process before the scan began.

A multilayer excitation echo-planar imaging (EPI) sequence was used to capture the resting-state fMRI images in the horizontal plane. The following parameters were applied to the sequence: repetition time (TR) =800 ms, echo time (TE) =37 ms, inclination angle =52°, matrix size =104×104, field of view (FOV) =208×208 mm², no slice gap, and voxel size =2×2×2 mm³. The participants were reminded to lie calmly in the scanning device, close their eyes without falling asleep, and maintain as little head movement as possible while scanning the resting-state functional images. The resting-state functional scanning procedure was completed in 404 seconds, and each subject received 488 functional images (one image every 0.8 seconds).

T1-weighted images were acquired using a magnetization-prepared rapid gradient-echo (MP-RAGE) sequence in the sagittal plane. The specific parameters were as follows: matrix size =320×320, FOV =256×256 mm², layer thickness =0.8 mm, voxel size =0.8×0.8×0.8 mm³, TR =3,000 ms, TE =2.56 ms, inversion time (TI) =1,100 ms, inclination angle =7°, and number of layers =208. Representative preprocessed fMRI time series and connectivity matrices are illustrated in Figure 1 [raw blood oxygenation level-dependent (BOLD) signal (Figure 1A); motion-corrected data (Figure 1B)].

Figure 1 Comparison of sFC analysis across different groups. (A) sFC analysis between the HC and 7-day groups. Notable alterations in connectivity of HIP R, PHG R, PUT L, and PAL L. Connection strength between PUT L and PAL L was weaker in the 7-day group versus the HCs group (P<0.05, FDR-corrected). (B) sFC analysis between the HC and 3-month groups. The sFC analysis between HCs and 3-month groups (P<0.001, uncorrected) showed changes in HIP-L, PHG L, PUT L-PAL L connections. The analysis also revealed weakened connections between CAU R, PAL R, OLF R, and TPOsup R (P<0.001, uncorrected). The color scale represents the strength of the t-value. CAU R, right caudate; FDR, false discovery rate; HC, healthy control; HIP L, left hippocampus; HIP R, right hippocampus; OLF R, right olfactory cortex; PAL L, left pallidum; PAL R, right pallidum; PHG L, left para-hippocampal gyrus; PHG R, right para-hippocampal gyrus; PUT L, left putamen; sFC, static functional connectivity; TPOsup R, right superior temporal pole.

Image preprocessing

All resting-state data were processed using the fMRI data preprocessing toolkit GRETNA (version 2.0, http://www.nitrc.org/projects/gretna/) based on the MATLAB platform to preprocess the data for analysis. The procedural steps were defined as follows: (I) conversion of format from Digital Imaging and Communications in Medicine (DICOM) to Neuroimaging Informatics Technology Initiative (NIfTI); (II) exclusion of the initial 10-time points to reduce changes in magnetic field stability during scanning and additional noise from external machinery, as well as to facilitate subject adaptation to the scanning environment, with the subsequent 478-time points utilized for further analyses; (III) time-layer calibration and head-motion correction; (IV) alignment and spatial standardization using high-resolution T1 image alignment and spatial normalization; (V) regress-out variables, including Friston 24 motion parameters, white matter signals, and cerebrospinal fluid (CSF) signals; (VI) linear regression and temporal band-pass filtering (0.01–0.1 Hz) were conducted to mitigate high-frequency noise (36-40).

sFC

Pre-processed resting-state fMRI data were used to construct the functional networks for each participant. The traditional resting-state FC method assumes that functional interactions are smooth in space and time, and the sFC is defined as a measure of the average connectivity over the entire scanning period (41). To separate the whole brain into 90 regions, we employed automatic anatomical labelling (AAL) mapping (42) to split the BOLD time-series signals into their respective 90 ROIs. FC is represented as a matrix, where rows and columns denote nodes, and each matrix element denotes the edge strength or FC between the corresponding nodes. Pearson’s correlation coefficient is the predominant linear approach for evaluating FC and is frequently employed to measure the synchronization of two signals. The FC between each pair of regions was defined as an edge, and the intensity of the FC between two ROIs was determined using Pearson’s correlation coefficient between the BOLD signals. In addition, to improve the normality of the data distribution, Fisher’s z-transformation was applied to the FC matrix values to generate the final 90×90 connectivity matrix, which characterizes the sFC matrix for each participant (43,44).

The Pearson correlation coefficient (Pearson correlation) is a statistical measure of the degree to which different parts of the brain can spontaneously coordinate their activity using fMRI. The greater the absolute value of the correlation coefficient obtained from the Pearson correlation analysis, the stronger the synchronization of activity between the two brain regions, and thus, the stronger the connection and connectivity between them.

dFC

The complete procedure for the dFC analysis was performed using a sliding window methodology. 90 ROIs were generated using the AAL maps. These 90 ROIs were classified into seven functional networks based on previous studies: the visual system (VIS), somatomotor network (SMN), dorsal attention network (DAN), ventral attention network (VAN), limbic network (LIM), frontoparietal network (FPN), and DMN (45,46). The VIS network within the brain is crucial for interpreting visual stimuli and facilitating actions based on this sensory input. An IS affecting the VIS network can cause visual field defects and complex issues like object agnosia (47). The SMN is a vital neural system within the brain, playing a key role in the formulation, performance, and coordination of voluntary motor activities. A stroke can severely impact the SMN, causing motor impairments like paralysis, balance issues, and sensory loss (48). A variety of cognitive and attentional deficits may be the consequence of impairments in both the DAN and the VAN in cases of IS. The VAN’s primary responsibility is to detect and respond to critical, unexpected, or behaviorally significant events in the environment (49). The Lateral Medullary Nucleus (LMN) is crucial in the context of various psychiatric and neurological disorders due to its profound involvement in processes related to feeling, memory, and inspiration (50). The FPN participates in multiple cognitive procedures, and disorders within this network has been linked to a variety of psychological, cognitive, and mental health conditions (51). The DMN is a group of brain areas that show increased activity when a person is at rest and not focused on external environments or specific tasks (52). The dFC splits a subject’s time points into time windows of a particular length and measures the correlation between the time sequences. An important parameter of the stationary state dFC is the window length. The sliding window method uses the window length as a key parameter. The minimum window length should not be less than 1/fmin (fmin represents the minimum frequency of the time course). The dynamic characteristics of the brain are more accurately represented by a window length of 30–50 TR, as demonstrated in previous studies (53). After preprocessing, the data had 478 temporal nodes; therefore, we chose a window length of 50 TR to balance the fast-changing dynamic relationships and reliable correlation estimation between the regions. We computed 429 individual windows in steps of 1 TR (with the window overlap configured to 98%). Pearson correlations between the time series from all ROI pairings were calculated within the specified windows. The overall ROI pair dynamic FC values were the standard deviation (SD) of Pearson’s correlations throughout all windows. A brain-wide temporal dynamics dFC matrix was created using these values. For each participant, Fisher’s r-to-z transformation was applied to the dFC matrix in all windows of the target dFC state to improve the normal distribution of the correlations. The final dynamic functional network matrix dimensions for each participant were 429×90×90 pixels in size.

After the computation was completed, cluster analysis was performed on all Fisher Z-transformed dFC matrices using the dynamic brain connectome (DynamicBC) Toolbox (54). To identify recurring FC states, k-means clustering was performed on the 429 windowed dFC matrices from the previous analytical stage. The distance between different brain FC patterns was measured using the squared Euclidean distance (SqEuclidean). The Euclidean distance, also known as the L2 norm, is the most common distance metric and measures the absolute distance between two points in a multidimensional space. The maximum number of states was configured to 20, whereas the optimal clustering centers were determined to be within the range of 2 to 20. Subsequently, the DynamicBC Toolbox automatically estimated the optimal number of states. The toolbox calculates the optimal number of states using the Silhouette, Calinski-Harabasz, and Davies-Bouldin metrics. The silhouette, which ranges from −1 to 1, reflects how similar a node is to its clusters compared with its neighbors. Larger numbers suggest that the node compares favorably; The Calinski-Harabasz index quantifies how similar a node is to the clusters it is part of to its neighboring nodes. A higher intra-cluster variance to inter-cluster variance ratio suggests a better clustering outcome in the Calinski-Harabasz index. The smaller the intra-cluster tightness and inter-cluster spacing, the better the Davies-Bouldin clustering. When the optimal value is considered, the final optimal state number is the average of the K values of all three metrics.

The following three temporal attribute features were used to characterize the temporal trajectories of the FC states observed in patients with AIS and HC—(I) mean dwell time (MDT): the average duration of consecutive states that are active, i.e., the MDT of an FC state is the average number of consecutive cycles spent in that state during the entire scanning period; (II) frequency of occurrence (F): the percentage of time points at which a state occurs; (III) Number of Transitions (NT): the probability of the number of transitions between functional states per second (55,56).

ML

This study involved the development of five distinct ML algorithm models: a linear-SVM, RBF-SVM (57), KNN (58), RF (59), and TREE (60). A linear-SVM is a supervised algorithm that categorizes data into two classes via a hyperplane or straight line. It reduces fMRI data complexity by focusing on significant features like brain regions or time points (61,62). RBF-SVM transforms data into higher dimensions to achieve linear separability. Complex and non-linear brain activation patterns have been found to be detected in fMRI. RBF-SVM uses a non-linear kernel to capture these non-linear correlations (63). For supervised tasks like regression and classification, the KNN algorithm is a great non-parametric learning option. In fMRI studies, KNN demonstrates effective adaptability to variations in brain activity patterns across different subjects or stimuli (64). RF is a vital ensemble learning method for regression and classification tasks because it is both simple and effective. It is engaged in the image obtaining process in fMRI (65). A TREE is a branch-structured predictive model that uses questions. The nodes represent feature-based decisions, while the leaves represent anticipated outcomes. These methods are used in data analysis to categorize or predict brain activity using fMRI data (66).

We used a grid search approach to optimize the model hyperparameters during the training process. The fundamental concept of the lattice search algorithm is to optimize the model performance by systematically traversing the given parameter combinations through a comprehensive search process. This approach involves conducting a lattice search for each combination in the list of hyperparameter combinations, instantiating the given model, performing K-fold cross-validation, and identifying the hyperparameter combinations that yield the highest average scores as the optimal choices for returning the model objects.

This study differs from prior research in that it utilizes features that exhibit significant differences between groups, identified through a two-sample t-test (P<0.01), rather than employing all-connection matrices as feature inputs, thereby excluding the irrelevant data. It has been reported in the literature that the use of all connections as features with redundant information may impair the classification results. Therefore, this study used support vector machine recursive feature elimination (SVM-RFE) as a strategy for feature selection. The optimal subset of features was filtered using the model. Each classifier underwent iterative training and evaluation using five-fold cross-validation (5-fold CV) and leave-one-out cross-validation (LOOCV) to mitigate overfitting by increasing the number of folds in the data. To measure the performance of the classifiers, we used the traditional definitions of recall (sensitivity), precision (specificity), F1 score (precision), and accuracy rate (accuracy) (67).

Group analysis

Statistical analyses of the differences between groups were conducted using a two-tailed two-sample t-test to examine variations between HCs and patients (7 days, 3 months) of different short durations, along with paired-sample t-tests among patients. The significance threshold p was set at 0.05, and multiple comparison corrections were performed using the false discovery rate (FDR) correction.

Experiment section

This research analyzed the cognitive performance and clinical characteristics of patients with AIS and HCs. Stroke patients were assessed in two groups at different times: one group at 7 days post-stroke (n=12) and the other at 3 months post-stroke (n=12). A control group of healthy individuals (n=26) was incorporated to establish a baseline for comparison purposes. Additionally, the clinical evaluations included three main variables: NHISS, MoCA, and ICARS. FC was assessed using fMRI, which tracks brain activity during a resting state in participants. Five distinct ML algorithms—linear-SVM, RBF-SVM, RF, KNN, and TREE—were utilized for feature selection and classification based on data obtained from brain connectivity studies. The sliding-window technique is utilized to examine the dynamic connectivity among brain regions. In the context of ML parameters, the SVM-RFE method is employed for feature selection, with the objective of identifying the most discriminative features for classification purposes. The robustness of the ML models was evaluated through 5-fold and LOOCV methods. Precision, recall, F1 score, and accuracy are vital metrics for evaluating the performance of ML classification models. The area under the curve (AUC) serves as an essential metric for assessing the classification performance of models, where higher AUC values indicate enhanced model efficacy.

Results

Demographic characteristics

Table 1 presents a full overview of the demographic and clinical data of the study, which encompasses stroke patients assessed at two distinct intervals—7 days and 3 months following the post-stroke event—as well as a group of HCs. The chart presents a comparative analysis of clinical and demographic data across the three cohorts [7 days group (n=12), 3 months group (n=12), and HCs group (n=26)]. The data present the distribution of sex, age, and clinical scores (NHISS, MoCA, and ICARS), accompanied by P values indicating statistical comparisons between the groups. Clinically, the NHISS (stroke severity) showed considerable improvement from 7 days (1.50±1.51) to 3 months (0.67±0.98), but still differed from that of HCs (P=0.12). Cognitive function, as assessed by the MoCA, demonstrated notable impairment at 7 days (19.42±10.64) in contrast to HCs (26.92±2.19), with only marginal improvement observed at 3 months (20.58±4.89), and no significant difference identified between the 7-day and 3-month assessments. Ataxia, evaluated using ICARS, is markedly more severe at 7 days (20±9.56) than in HCs (1.08±1.06), and exhibits improvement at 3 months (15.67±7.94), however, it remains significantly elevated compared to HCs, with no notable change between 7 days and 3 months. The overall results indicated a slight improvement in stroke severity, cognitive function, and ataxia after three months; nonetheless, patients still showed noticeable differences compared with healthy individuals.

sFC differences

Differential sFC analysis between the HCs and 7-day groups (P<0.05, FDR-corrected) revealed significant changes in the connectivity of the right hippocampus (HIP R), right para-hippocampal gyrus (PHG R), and left putamen (PUT L)-left pallidum (PAL L). Specifically, the connection strength between PUT L and PAL L was weaker in the 7-day group than in the HCs group (Figure 1A). Significant changes were observed in the DMN (HIP R, PHG R), and VAN (PUT L, PAL L). The sFC analysis between the HCs and 3-month groups (P<0.001, uncorrected) revealed significant changes in the left hippocampus (HIP L)-left para-hippocampal gyrus (PHG L) and PUT L-PAL L connections. In addition, the sFC analysis between the HCs and the 3-month group (P<0.001, uncorrected) revealed weakened connection strength between the right caudate (CAU R)-right pallidum (PAL R) and right olfactory cortex (OLF R)-right superior temporal pole (TPOsup R) (Figure 1B). Notable alterations were detected in the DMN (HIP L and PHG L), VAN (PUT L and PAL L), and LIM (CAU R and TPOsup R) networks. No significant differences were observed between the groups at 7 and 3 months.

dFC differences

We constructed a dFC brain network using a sliding-window approach. Subsequently, we analyzed and revealed two distinct patterns of dFC, namely State 1 and State 2, utilizing k-means clustering, and examined the temporal attribute characteristics of the dFC alongside its temporal variability characteristics. The clustering of the Euclidean distance matrix was performed using the DynamicBC Tool. The optimal number of states, as estimated by the Silhouette, Calinski-Harabasz, and Davies-Bouldin values, was K=2. Consequently, the FC matrix for each subject was categorized into one of two FC states. K-means clustering analysis of the results across all windows identified the center of mass, and the overall percentage of windows was allocated to the respective cluster (as shown above for each matrix). The overall frequency of occurrence of State 1 was 63.61% (Figure 2A), with matrix connections exhibiting a predominant pattern of weak connections. The frequency of occurrence of State 2 was 36.39% (Figure 2B), and the matrix connections predominantly exhibited strong connections. Consequently, we designated State 1 as the “high-frequency weakly connected” and State 2 is designated as the “low-frequency strong connection” state.

Figure 2 Clusters of center-of-mass maps for different states. (A) State 1 clustered center-of-mass map. (B) State 2 clustered center-of-mass map. The percentages (36.01% and 36.39%) represent the lesion volume relative to total brain volume for each patient case.

Figure 3 displays the outcomes of the between-group analysis of the temporal features across the three groups: 3 months, 7 days, and the HCs group. The fractional occupancy (F) and MDT of States 1 and 2 were not significantly different between the groups. The Number of Transitions (NT) exhibited a significant between-group difference in the statistical analysis between the HCs and 7 days group, whereas there was no significant between-group difference in the other two groups (Figure 3A). Patients in the experimental group were more likely to exhibit state switching after 7 days of therapy. MDT State 2 exhibited a negative correlation with MoCA, whereas NT demonstrated a positive correlation with MoCA (Figure 3B) and a negative correlation with ICARS (Figure 3C). The correlation between the MDT state and mini-mental state examination (MMSE) was evaluated (Figure 3D). Evaluation of the connection between MDT State 2 and MoCA (Figure 3E). Correlation analysis indicated a positive correlation between State 1 and MoCA scores, whereas State 2 showed a negative correlation with MoCA scores. A prolonged duration in State 1, characterized by high-frequency weak connections, correlates with increased symptom; State 1 is associated with cognitive impairment. A prolonged duration in the low-frequency strongly connected State 2 correlates with milder symptoms, indicating that State 2 is associated with enhanced cognitive flexibility. Post-treatment, transitioning to State 1 becomes more feasible.

Figure 3 Neuropsychological and motor assessment analyses. (A) Results of the analysis of differences between NT groups. (B) NT and MoCA correlation analysis. (C) NT and ICARS correlation analysis. (D) MDT and MMSE correlation analysis. (E) MDT State 2 and MoCA correlation analysis. *, P<0.05; HC, healthy control; ns, not significant. ICARS, International Cooperative Ataxia Rating Scale; MDT, mean dwell time; MMSE, mini-mental state examination; MoCA, Montreal Cognitive Assessment; NT, Number of Transitions.

Following the cluster analysis, each participant generated a unique center-of-mass matrix, represented by the participant’s median matrix. To analyze the within-state differences between the two dFC patterns, we performed a between-group difference analysis of the median matrices of the participants belonging to each state with the HCs group to obtain the differential change in within-state connectivity, that is, the strength of between-group connectivity. Statistical analyses of the HC versus the 7 days group showed (P<0.001, uncorrected) that in State 1 (Figure 4A), the left inferior parietal lobule (IPL L) and left Heschl’s gyrus (HES L), HIP R and PHG R, PAL L, and PUT L were attenuated, and the remaining connections were enhanced. Statistical analyses of the HCs versus the 3 months group showed (P<0.001, uncorrected) that all connections in State 1 were enhanced (Figure 4B). In State 2 (Figure 4C), right rectus (REC R) was enhanced with right lingual gyrus (LING R), right insula (INS R) was enhanced with left thalamus (THA L), INS R was enhanced with PHG L, left Rolandic operculum (ROL L) was enhanced with right angular gyrus (ANG R), and PHG L was weakened with right inferior parietal lobule (IPL R). In State 2 (Figure 4D), all connections in State 1 were attenuated.

Figure 4 Between-group dFC results. (A) dFC State 1 results for HC vs. 7 days (P<0.001, uncorrected). (B) dFC State1 results for HC vs. 3 months (P<0.001, uncorrected). (C) dFC State 2 results for 7 days vs. HCs (P<0.001, uncorrected). (D) dFC State 2 results for 3 months vs. HCs (P<0.001, uncorrected). The colors indicate the t-values, with red indicating enhanced t-values and blue indicating attenuated t-values. DAN, dorsal attention network; dFC, dynamic functional connectivity; DMN, default mode network; FPN, frontoparietal network; HC, healthy control; LIM, limbic network; SMN, somatomotor network; VAN, ventral attention network; VIS, visual system.

The results of the statistical analysis comparing the groups in terms of dFC variance are presented in (Figure 5). Connections within the DMN, LIM, and VIS were more likely to show significant variations. The results of the between-group dFC variance for HCs and 7 days [P<0.05, network-based statistic (NBS)-corrected] showed that all connections underwent significantly enhanced fluctuations in variability over time (Figure 5A). The results of the between-group dFC variance for HCs and 3 months (P<0.05, NBS-corrected) showed significant fluctuations in the variability of all connections, with an enhancement over time (Figure 5B). The results of the between-group dFC variance for 7 days and 3 months (P<0.001, uncorrected) also showed similar results (Figure 5C).

Figure 5 Between-group dFC variance results. (A) dFC variance results for HC vs. 7 days (P<0.05, NBS-corrected). (B) dFC variance results for HC vs. 3 months (P<0.05, NBS-corrected). (C) dFC variance results for 7 days vs. 3 months (P<0.001, uncorrected). The lines indicate the t-values, with red indicating enhanced t-values and cyan indicating attenuated t-values. Color key: The seven colors represent the seven functional networks (e.g., Default Mode Network, Frontoparietal Network, etc.) used in the analysis. dFC, dynamic functional connectivity; HC, healthy control; NBS, network-based statistic.

ML model

Initially, we identified 289 features that exhibited significant differences (two-sample t-test, P<0.01) between the HCs and patient groups, to exclude features with redundant information. Utilizing the inter-group differences feature selection method and SVM-RFE for specialized screening, the RBF-SVM identifies 38 optimal feature subsets (Figure 6A), linear-SVM identifies 1 optimal feature subset (Figure 6B), RF identifies 5 optimal feature subsets (Figure 6C), KNN identifies 74 optimal feature subsets (Figure 6D), and TREE identifies 174 optimal feature subsets (Figure 6E). The outcomes of 5-fold CV and LOOCV following the acquisition of optimal feature subsets through the five ML models via the feature selection process are presented in Table 2. This table summarizes the recall, precision, F1 scores, and classification accuracies for the various models. Regarding the AUC, the RBF-SVM model exhibited superior performance, followed by RF and TREE. The TREE model exhibited the highest accuracy, followed by the RBF-SVM and linear-SVM models, respectively. Five ML algorithms were implemented. (I) Linear-SVM: linear decision boundaries; (II) RBF-SVM: non-linear kernel for complex patterns; (III) KNN: distance-based classification; (IV) RF: ensemble feature selection; (V) TREE: interpretable hierarchical rules.

Figure 6 Optimal feature selection for different classifiers. (A) Optimal features of the RBF-SVM. (B) Optimal features of the linear SVM. (C) Optimal features of the RF. (D) Optimal features of the KNN. (E) Optimal features of the TREE. DMN, default mode network; KNN, K-nearest neighbors; LIM, limbic network; RBF-SVM, radial basis function support vector machine; RF, random forest; SVM, support vector machine; TREE, decision tree; VAN, ventral attention network; VIS, visual system.

Table 3 illustrates the alterations in the ROI across each functional network as each model was evaluated to identify the optimal feature subset during feature selection. The optimal feature subset indicates that, except for the linear-SVM model, the remaining models have the highest number of altered nodes in the VIS, which has undergone substantial changes, followed by DMN and LIM, which have also changed more significantly. The significance of feature weights in classification verified our previous findings, indicating substantial alterations in the VIS, DMN, and LIM networks. The comparison of the outcomes obtained from the three distinct approaches sFC, dFC, and ML is presented in Table 4.

Discussion

In this study, we initially conducted a whole-brain sFC analysis and discovered that the VIS and DMN were significantly modified in individuals with AIS. Next, we performed a sliding window dFC analysis on State 1 of the connectivity pattern, which is largely high-frequency weak connectivity, and State 2, which is predominantly low-frequency strong connectivity. The dFC analysis revealed significant changes in the connectivity of the DMN, VIS, FPN, and LIM networks. Finally, we established five ML models based on the sFC features of the intergroup differences: linear-SVM, SVM-RBF, KNN, RF, and TREE. By optimizing the AUC, the best feature subset showed that the SFC in the VIS, DMN, and LIM classified patients with AIS and HCs. AIS neural processes may be revealed by FC changes in the VIS, DMN, and LIM networks of the brain.

sFC provides efficient whole-brain connectivity mapping but lacks temporal dynamics; dFC captures transient states at a higher computational cost; and ML enables predictive modeling but requires larger samples. Clinical translation: sFC for initial screening, dFC for monitoring acute recovery, and ML for long-term prognosis using cloud-based processing (24).

The sFC analysis showed significant changes in the VAN, DMN, and LIM networks, which may have improved FC after treatment. Prior research has demonstrated atypical alterations in functional brain networks in patients with AIS (68,69). This study identified significant alterations in the VAN and DMN in the sFC at both 7 days and 3 months, as well as notable changes in the bean-shaped shell nucleus (PUT L) and bean-shaped pallidum (PAL R) at the 3-month point. The basal ganglia primarily consist of the bean-shaped shell nucleus and pallidum (70,71). Giroud et al. identified two clinical syndromes linked to lesions localized in the nucleus of the crystalline lens (72): (I) behavioral and cognitive deficits related to infarction in the pallidum; and (II) dyskinesia (dystonia) alongside cognitive deficits associated with shell nucleus disorders. cognitive deficits. A study of the bean-shaped shell nucleus and pallidum revealed that chronic disorders can cause Parkinson’s syndrome and other movement disorders, possibly leading to a stroke. An analysis of FC in patients with chronic-phase occipital stroke exhibiting ipsilateral visual field deficits before and after repetitive transorbital alternating current stimulation (rtACS), indicated significant alterations in ROIs within the occipital lobe post-treatment (73), and the reduction in FC between visual areas showed significant improvement over a 3-month period, consistent with findings from prior research (74). Furthermore, hyperarousal of the DMN and hypersensitivity to audiovisual stimuli have been identified as potential primary mechanisms of insomnia in patients who have experienced IS, contributing to cognitive decline and impaired emotional regulation (75), and impairment of DMN function has been observed in patients with altered consciousness due to various etiologies (76). The DMN functions within a network of dynamic interactions that influence its operation, and stroke disrupts the intrinsic FC of the DMN and its relationship with other networks (77,78). A greater degree of connectivity between the VAN and the DMN was correlated with a larger degree of patient improvement on cognitive tests in visual learning tests within 6 months of IS (75). During recovery, DMN connectivity is re-established in most patients with stroke, and a decrease in DMN activity may indicate permanent cognitive effects. Another study found disrupted FC in the DMN after stroke, causing cognitive and affective abnormalities in patients with stroke. DMN disruption can cause additional comorbid diseases, indicating that anxiety and depression in the first month after stroke are related to greater FC.

Following k-means cluster analysis of the dFC matrices of 54 subjects based on the Silhouette, Calinski-Harabasz, and Davies-Bouldin values, we identified two states of dFC that were strongly associated with symptom severity. State 1 exhibited a connectivity pattern characterized by high-frequency weak connections, whereas State 2 exhibited a connectivity pattern characterized by low-frequency strong connections. Patients with stroke and severe impairment demonstrated a significant increase in the duration spent in State 1. The patients exhibited a significant increase in state switches, a higher probability of transitioning to the robust connection mode of State 2, more frequent occurrences of State 2, and prolonged duration in this condition after three months. In addition, symptom improvement was strongly associated with time spent in State 2. These links pairings varied substantially between the stroke patient groups in both states in the correlation analysis using the measures. State 1 was negatively correlated with MoCA scores, whereas State 2 was positively correlated with MoCA scores. Moreover, the positive influence of State 2 on treatment efficacy was further verified by the observation that an increased frequency of State 2 occurrences correlated with improved treatment outcomes in patients with stroke. In addition, we detected widespread differences in dFC between brain regions with different functional domains. The analysis of between-group differences in dFC revealed that alterations in the State 1 network were primarily influenced by changes in the DMN, VIS, and LIM, consistent with prior research findings (79).

In a thalamic stroke study, cognitive domains were linked to the frontal-parietal-cerebellar-thalamic loops, whereas language areas were linked to the supplementary motor area (SMA), inferior frontal gyrus (IFG), and language-related brain regions. The bilateral SMA is likely crucial for speech recovery, while right language-related regions, such as the IFG, angular gyrus, and supramarginal gyrus, may influence the recovery from thalamic aphasia (80). The components of working memory are associated with a network of brain regions, notably the left angular gyrus and the posterior frontal cortex (81). A multivariate lesion-symptom map was used to identify strategic brain regions for post-stroke cognitive deficits in a large cohort of 410 patients with AIS. The MoCA was administered 3–6 months after the stroke to assess overall cognitive functioning and the cognitive domains. Support vector regression was used in multivariate analyses at the voxel and region of interest levels to assess cognition and infarct site. The left angular gyrus, basal ganglia structures, and peri basal ganglia white matter were critical structures for stroke-related cognitive impairment in both hypothesis-free analyses. A strategic network involving several overlapping and domain-specific cortical and subcortical structures has been identified for each cognitive domain (82). These findings confirm that stroke could cause angular gyrus injury and that motor impairment is linked with reduced thalamus, chiasmatic nucleus, pallidum, amygdala, and hippocampal sizes in children with AIS (83).

Finally, we employed an SVM-based ML approach to identify significant neuroimaging biomarkers for classifying patients with AIS and HCs. First, we identified 289 sFC-connected features that exhibited significant between-group differences using a two-sample t-test for the ML model. Second, feature weights were assigned, and the between-group difference features were ordered in descending order based on SVM-RFE to identify the most discriminative feature connections. In the classification stage, five supervised learning classifiers were employed: linear-SVM, RBF-SVM, RF, KNN, and TREE. We used multiple indicators to evaluate the classifier performance to ensure overall results and prevent errors (e.g., overemphasizing a statistic). This study confirmed our previous findings that the VIS changed the most within the subset of optimal connectivity parameters, followed by the LIM and the DMN. The VIS changed drastically, helping ML and FC categorization.

There are several limitations in this study. First, the limited sample size restricts comparative analyses between the groups; therefore, future studies should involve larger cohorts of patients with unilateral AIS. Second, this study involved patients with IS and stroke-induced impairment of the central visual pathway within three months of stroke onset; thus, further investigation is necessary to understand the long-term dynamic patterns. Finally, we divided the entire brain into 90 regions based on previously published maps to create functional brain networks. Additional research is required to identify the optimal brain-wrapping strategy or spatial scale for characterizing connections, as varying wrapping schemes and spatial scales demonstrate distinct topologies. Larger sample sizes and more detailed brain maps are essential for developing more discriminative ML classification models.

Conclusions

This study summarizes the results of an analysis using the AAL 90 atlas. First, the results showed that the VIS, DMN, and LIM networks were significantly altered in patients with AIS according to the sFC and dFC analyses. Second, ML models designed to differentiate patients with HCs and AIS indicated that the connectivity features of the three networks exhibited comparable significance in terms of their exclusive strengths. Future studies on the rehabilitation and treatment of patients with stroke may focus on abnormal brain regions, which may also shed light on the neurological processes that cause AIS.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-834/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-834/dss

Funding: This work was supported by Shanghai Science and Technology Innovation Action Plan (No. 23DZ2291500).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-834/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics Committee of the Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (approval number: 2021-121). All the study participants or their designated representatives provided written informed consent. Notably, individuals experiencing cognitive decline may lack the capacity to autonomously consent to participating in research studies. On their behalf, a legal guardian or representative gave informed consent for participation.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Kessner SS, Bingel U, Thomalla G. Somatosensory deficits after stroke: a scoping review. Top Stroke Rehabil 2016;23:136-46. [Crossref] [PubMed]
2. Van Kaam RC, van Putten MJAM, Vermeer SE, Hofmeijer J. Contralesional Brain Activity in Acute Ischemic Stroke. Cerebrovasc Dis 2018;45:85-92. [Crossref] [PubMed]
3. Feske SK. Ischemic Stroke. Am J Med 2021;134:1457-64. [Crossref] [PubMed]
4. Paul S, Candelario-Jalil E. Emerging neuroprotective strategies for the treatment of ischemic stroke: An overview of clinical and preclinical studies. Exp Neurol 2021;335:113518. [Crossref] [PubMed]
5. Ekkert A, Šliachtenko A, Grigaitė J, Burnytė B, Utkus A, Jatužis D. Ischemic Stroke Genetics: What Is New and How to Apply It in Clinical Practice? Genes (Basel) 2021;13:48. [Crossref] [PubMed]
6. Walter K. What Is Acute Ischemic Stroke? JAMA 2022;327:885. [Crossref] [PubMed]
7. Marto JP, Strambo D, Livio F, Michel P. Drugs Associated With Ischemic Stroke: A Review for Clinicians. Stroke 2021;52:e646-59. [Crossref] [PubMed]
8. Fox MD, Corbetta M, Snyder AZ, Vincent JL, Raichle ME. Spontaneous neuronal activity distinguishes human dorsal and ventral attention systems. Proc Natl Acad Sci U S A 2006;103:10046-51. [Crossref] [PubMed]
9. Deshpande G, Wang P, Rangaprakash D, Wilamowski B. Fully Connected Cascade Artificial Neural Network Architecture for Attention Deficit Hyperactivity Disorder Classification From Functional Magnetic Resonance Imaging Data. IEEE Trans Cybern 2015;45:2668-79. [Crossref] [PubMed]
10. Jang H, Plis SM, Calhoun VD, Lee JH. Task-specific feature extraction and classification of fMRI volumes using a deep neural network initialized with a deep belief network: Evaluation using sensorimotor tasks. Neuroimage 2017;145:314-28. [Crossref] [PubMed]
11. Laumann TO, Snyder AZ, Mitra A, Gordon EM, Gratton C, Adeyemo B, Gilmore AW, Nelson SM, Berg JJ, Greene DJ, McCarthy JE, Tagliazucchi E, Laufs H, Schlaggar BL, Dosenbach NUF, Petersen SE. On the Stability of BOLD fMRI Correlations. Cereb Cortex 2017;27:4719-32. [Crossref] [PubMed]
12. Liu J, Qin W, Zhang J, Zhang X, Yu C. Enhanced interhemispheric functional connectivity compensates for anatomical connection damages in subcortical stroke. Stroke 2015;46:1045-51. [Crossref] [PubMed]
13. Xu H, Qin W, Chen H, Jiang L, Li K, Yu C. Contribution of the resting-state functional connectivity of the contralesional primary sensorimotor cortex to motor recovery after subcortical stroke. PLoS One 2014;9:e84729. [Crossref] [PubMed]
14. Carter AR, Astafiev SV, Lang CE, Connor LT, Rengachary J, Strube MJ, Pope DL, Shulman GL, Corbetta M. Resting interhemispheric functional magnetic resonance imaging connectivity predicts performance after stroke. Ann Neurol 2010;67:365-75. [Crossref] [PubMed]
15. Park CH, Chang WH, Ohn SH, Kim ST, Bang OY, Pascual-Leone A, Kim YH. Longitudinal changes of resting-state functional connectivity during motor recovery after stroke. Stroke 2011;42:1357-62. [Crossref] [PubMed]
16. Xie H, Li X, Huang W, Yin J, Luo C, Li Z, Dou Z. Effects of robot-assisted task-oriented upper limb motor training on neuroplasticity in stroke patients with different degrees of motor dysfunction: A neuroimaging motor evaluation index. Front Neurosci 2022;16:957972. [Crossref] [PubMed]
17. Huo C, Sun Z, Xu G, Li X, Xie H, Song Y, Li Z, Wang Y. fNIRS-based brain functional response to robot-assisted training for upper-limb in stroke patients with hemiplegia. Front Aging Neurosci 2022;14:1060734. [Crossref] [PubMed]
18. Nguyen VT, Lu YH, Wu CW, Sung PS, Lin CC, Lin PY, Wang SM, Chen FY, Chen JJ. Evaluating interhemispheric synchronization and cortical activity in acute stroke patients using optical hemodynamic oscillations. J Neural Eng 2022;
19. Liu Q, Wang B, Liu Y, Lv Z, Li W, Li Z, Fan Y. Frequency-specific Effective Connectivity in Subjects with Cerebral Infarction as Revealed by NIRS Method. Neuroscience 2018;373:169-81. [Crossref] [PubMed]
20. Magalhães R, Picó-Pérez M, Esteves M, Vieira R, Castanho TC, Amorim L, Sousa M, Coelho A, Fernandes HM, Cabral J, Moreira PS, Sousa N. Habitual coffee drinkers display a distinct pattern of brain functional connectivity. Mol Psychiatry 2021;26:6589-98. [Crossref] [PubMed]
21. Sen B, Cullen KR, Parhi KK. Classification of Adolescent Major Depressive Disorder Via Static and Dynamic Connectivity. IEEE J Biomed Health Inform 2021;25:2604-14. [Crossref] [PubMed]
22. Liu F, Wang Y, Li M, Wang W, Li R, Zhang Z, Lu G, Chen H. Dynamic functional network connectivity in idiopathic generalized epilepsy with generalized tonic-clonic seizure. Hum Brain Mapp 2017;38:957-73. [Crossref] [PubMed]
23. Preti MG, Bolton TA, Van De Ville D. The dynamic functional connectome: State-of-the-art and perspectives. Neuroimage 2017;160:41-54. [Crossref] [PubMed]
24. Pervaiz U, Vidaurre D, Woolrich MW, Smith SM. Optimising network modelling methods for fMRI. Neuroimage 2020;211:116604. [Crossref] [PubMed]
25. Tijhuis FB, Broeders TAA, Santos FAN, Schoonheim MM, Killestein J, Leurs CE, van Geest Q, Steenwijk MD, Geurts JJG, Hulst HE, Douw L. Dynamic functional connectivity as a neural correlate of fatigue in multiple sclerosis. Neuroimage Clin 2021;29:102556. [Crossref] [PubMed]
26. Supekar K, Cai W, Krishnadas R, Palaniyappan L, Menon V. Dysregulated Brain Dynamics in a Triple-Network Saliency Model of Schizophrenia and Its Relation to Psychosis. Biol Psychiatry 2019;85:60-9. [Crossref] [PubMed]
27. Barttfeld P, Uhrig L, Sitt JD, Sigman M, Jarraya B, Dehaene S. Signature of consciousness in the dynamics of resting-state brain activity. Proc Natl Acad Sci U S A 2015;112:887-92. [Crossref] [PubMed]
28. Lurie DJ, Kessler D, Bassett DS, Betzel RF, Breakspear M, Kheilholz S, et al. Questions and controversies in the study of time-varying functional connectivity in resting fMRI. Netw Neurosci 2020;4:30-69. [Crossref] [PubMed]
29. Kamal H, Lopez V, Sheth SA. Machine Learning in Acute Ischemic Stroke Neuroimaging. Front Neurol 2018;9:945. [Crossref] [PubMed]
30. Hu X, Qi D, Li S, Ye S, Chen Y, Cao W, Du M, Zheng T, Li P, Fang Y. Development and Validation of an Interpretable Machine Learning Model for Prediction of the Risk of Clinically Ineffective Reperfusion in Patients Following Thrombectomy for Ischemic Stroke. Ther Clin Risk Manag 2025;21:621-31. [Crossref] [PubMed]
31. Teasdale G, Jennett B. Assessment of coma and impaired consciousness. A practical scale. Lancet 1974;2:81-4. [Crossref] [PubMed]
32. Trouillas P, Takayanagi T, Hallett M, Currier RD, Subramony SH, Wessel K, Bryer A, Diener HC, Massaquoi S, Gomez CM, Coutinho P, Ben Hamida M, Campanella G, Filla A, Schut L, Timann D, Honnorat J, Nighoghossian N, Manyam B. International Cooperative Ataxia Rating Scale for pharmacological assessment of the cerebellar syndrome. The Ataxia Neuropharmacology Committee of the World Federation of Neurology. J Neurol Sci 1997;145:205-11. [Crossref] [PubMed]
33. Kasner SE. Clinical interpretation and use of stroke scales. Lancet Neurol 2006;5:603-12. [Crossref] [PubMed]
34. Adams HP Jr, Davis PH, Leira EC, Chang KC, Bendixen BH, Clarke WR, Woolson RF, Hansen MD. Baseline NIH Stroke Scale score strongly predicts outcome after stroke: A report of the Trial of Org 10172 in Acute Stroke Treatment (TOAST). Neurology 1999;53:126-31. [Crossref] [PubMed]
35. Nasreddine ZS, Phillips NA, Bédirian V, Charbonneau S, Whitehead V, Collin I, Cummings JL, Chertkow H. The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment. J Am Geriatr Soc 2005;53:695-9. [Crossref] [PubMed]
36. Zheng Y, Wu Y, Liu Y, Li D, Liang X, Chen Y, Zhang H, Guo Y, Lu R, Wang J, Qiu S. Abnormal dynamic functional connectivity of thalamic subregions in patients with first-episode, drug-naïve major depressive disorder. Front Psychiatry 2023;14:1152332. [Crossref] [PubMed]
37. Liu Y, Li F, Shang S, Wang P, Yin X, Krishnan Muthaiah VP, Lu L, Chen YC. Functional-structural large-scale brain networks are correlated with neurocognitive impairment in acute mild traumatic brain injury. Quant Imaging Med Surg 2023;13:631-44. [Crossref] [PubMed]
38. Zhang Y, Xiang Q, Huang CC, Zhao J, Liu Y, Lin CP, Liu D, Lo CZ. Short-term Medication Effects on Brain Functional Activity and Network Architecture in First-Episode psychosis: a longitudinal fMRI study. Brain Imaging Behav 2023;17:137-48. [Crossref] [PubMed]
39. Yang L, Jin C, Qi S, Teng Y, Li C, Yao Y, Ruan X, Wei X. Aberrant degree centrality of functional brain networks in subclinical depression and major depressive disorder. Front Psychiatry 2023;14:1084443. [Crossref] [PubMed]
40. Lindquist MA, Geuter S, Wager TD, Caffo BS. Modular preprocessing pipelines can reintroduce artifacts into fMRI data. Hum Brain Mapp 2019;40:2358-76. [Crossref] [PubMed]
41. Schneider SC, Archila-Meléndez ME, Göttler J, Kaczmarz S, Zott B, Priller J, Kallmayer M, Zimmer C, Sorg C, Preibisch C. Resting-state BOLD functional connectivity depends on the heterogeneity of capillary transit times in the human brain A combined lesion and simulation study about the influence of blood flow response timing. Neuroimage 2022;255:119208. [Crossref] [PubMed]
42. Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F, Etard O, Delcroix N, Mazoyer B, Joliot M. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. Neuroimage 2002;15:273-89. [Crossref] [PubMed]
43. Eichenbaum A, Pappas I, Lurie D, Cohen JR, D'Esposito M. Differential contributions of static and time-varying functional connectivity to human behavior. Netw Neurosci 2021;5:145-65. [Crossref] [PubMed]
44. Lewis N, Lu H, Liu P, Hou X, Damaraju E, Iraji A, Calhoun V. Static and dynamic functional connectivity analysis of cerebrovascular reactivity: An fMRI study. Brain Behav 2020;10:e01516. [Crossref] [PubMed]
45. Jiang X, Hou C, Ma J, Li H. Alterations in local activity and whole-brain functional connectivity in human immunodeficiency virus-associated neurocognitive disorders: a resting-state functional magnetic resonance imaging study. Quant Imaging Med Surg 2025;15:563-80. [Crossref] [PubMed]
46. Jang H, Mashour GA, Hudetz AG, Huang Z. Measuring the dynamic balance of integration and segregation underlying consciousness, anesthesia, and sleep in humans. Nat Commun 2024;15:9164. [Crossref] [PubMed]
47. Van Orden K, Meyer DM, Perrinez ES, Torres D, Poynor B, Alwood B, Bykowski J, Khalessi A, Meyer BC. (VISIION-S): Viz.ai Implementation of Stroke augmented Intelligence and communications platform to improve Indicators and Outcomes for a comprehensive stroke center and Network - Sustainability. J Stroke Cerebrovasc Dis 2023;32:107303. [Crossref] [PubMed]
48. Zhang Z. Network Abnormalities in Ischemic Stroke: A Meta-analysis of Resting-State Functional Connectivity. Brain Topogr 2025;38:19. [Crossref] [PubMed]
49. Adhikari MH, Griffis J, Siegel JS, Thiebaut de Schotten M, Deco G, Instabato A, Gilson M, Corbetta M. Effective connectivity extracts clinically relevant prognostic information from resting state activity in stroke. Brain Commun 2021;3:fcab233. [Crossref] [PubMed]
50. Li P, Chen C, Huang B, Jiang Z, Wei J, Zeng J. Altered excitability of motor neuron pathways after stroke: more than upper motor neuron impairments. Stroke Vasc Neurol 2022;7:518-26. [Crossref] [PubMed]
51. Olafson E, Russello G, Jamison KW, Liu H, Wang D, Bruss JE, Boes AD, Kuceyeski A. Frontoparietal network activation is associated with motor recovery in ischemic stroke patients. Commun Biol 2022;5:993. [Crossref] [PubMed]
52. Tuladhar AM, Snaphaan L, Shumskaya E, Rijpkema M, Fernandez G, Norris DG, de Leeuw FE. Default Mode Network Connectivity in Stroke Patients. PLoS One 2013;8:e66556. [Crossref] [PubMed]
53. Savva AD, Mitsis GD, Matsopoulos GK. Assessment of dynamic functional connectivity in resting-state fMRI using the sliding window technique. Brain Behav 2019;9:e01255. [Crossref] [PubMed]
54. Liao W, Wu GR, Xu Q, Ji GJ, Zhang Z, Zang YF, Lu G. DynamicBC: a MATLAB toolbox for dynamic brain connectome analysis. Brain Connect 2014;4:780-90. [Crossref] [PubMed]
55. Luo Q, Chen J, Li Y, Wu Z, Lin X, Yao J, Yu H, Wu H, Peng H. Aberrant static and dynamic functional connectivity of amygdala subregions in patients with major depressive disorder and childhood maltreatment. Neuroimage Clin 2022;36:103270. [Crossref] [PubMed]
56. Li Y, Ran Y, Chen Q. Abnormal static and dynamic functional network connectivity of the whole-brain in children with generalized tonic-clonic seizures. Front Neurosci 2023;17:1236696. [Crossref] [PubMed]
57. Ozaltin O, Coskun O, Yeniay O, Subasi A. A Deep Learning Approach for Detecting Stroke from Brain CT Images Using OzNet. Bioengineering (Basel) 2022;9:783. [Crossref] [PubMed]
58. Yun K, He T, Zhen S, Quan M, Yang X, Man D, Zhang S, Wang W, Han X. Development and validation of explainable machine-learning models for carotid atherosclerosis early screening. J Transl Med 2023;21:353. [Crossref] [PubMed]
59. Wang J, Gong X, Chen H, Zhong W, Chen Y, Zhou Y, Zhang W, He Y, Lou M. Causative Classification of Ischemic Stroke by the Machine Learning Algorithm Random Forests. Front Aging Neurosci 2022;14:788637. [Crossref] [PubMed]
60. Ajčević M, Miladinović A, Furlanis G, Buoite Stella A, Naccarato M, Caruso P, Manganotti P, Accardo A. Wake-up Stroke Outcome Prediction by Interpretable Decision Tree Model. Stud Health Technol Inform 2022;294:569-70. [Crossref] [PubMed]
61. Pereira F, Mitchell T, Botvinick M. Machine learning classifiers and fMRI: a tutorial overview. Neuroimage 2009;45:S199-209. [Crossref] [PubMed]
62. Richiardi J, Eryilmaz H, Schwartz S, Vuilleumier P, Van De Ville D. Decoding brain states from fMRI connectivity graphs. Neuroimage 2011;56:616-26. [Crossref] [PubMed]
63. Wang Z, Childress AR, Wang J, Detre JA. Support vector machine learning-based fMRI data group analysis. Neuroimage 2007;36:1139-51. [Crossref] [PubMed]
64. Yang Q, Zhang H, Xia J, Zhang X. Evaluation of magnetic resonance image segmentation in brain low-grade gliomas using support vector machine and convolutional neural network. Quant Imaging Med Surg 2021;11:300-16. [Crossref] [PubMed]
65. Gresser E, Schachtner B, Stüber AT, Solyanik O, Schreier A, Huber T, Froelich MF, Magistro G, Kretschmer A, Stief C, Ricke J, Ingrisch M, Nörenberg D. Performance variability of radiomics machine learning models for the detection of clinically significant prostate cancer in heterogeneous MRI datasets. Quant Imaging Med Surg 2022;12:4990-5003. [Crossref] [PubMed]
66. Li Z, Chen L, Song Y, Dai G, Duan L, Luo Y, Wang G, Xiao Q, Li G, Bai S. Predictive value of magnetic resonance imaging radiomics-based machine learning for disease progression in patients with high-grade glioma. Quant Imaging Med Surg 2023;13:224-36. [Crossref] [PubMed]
67. Ruksakulpiwat S, Thongking W, Zhou W, Benjasirisan C, Phianhasin L, Schiltz NK, Brahmbhatt S. Machine learning-based patient classification system for adults with stroke: A systematic review. Chronic Illn 2023;19:26-39. [Crossref] [PubMed]
68. Chai Y, Sheline YI, Oathes DJ, Balderston NL, Rao H, Yu M. Functional connectomics in depression: insights into therapies. Trends Cogn Sci 2023;27:814-32. [Crossref] [PubMed]
69. Zhang J, Chang Y. Alterations of static and dynamic functional network connectivity in acute ischemic brainstem stroke. Acta Radiol 2023;64:1623-30. [Crossref] [PubMed]
70. Javed N, Cascella M. Neuroanatomy, Globus Pallidus. In: StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing; 2025 Jan.
71. Chung SJ, Im JH, Lee MC, Kim JS. Hemichorea after stroke: clinical-radiological correlation. J Neurol 2004;251:725-9. [Crossref] [PubMed]
72. Giroud M, Lemesle M, Madinier G, Billiar T, Dumas R. Unilateral lenticular infarcts: radiological and clinical syndromes, aetiology, and prognosis. J Neurol Neurosurg Psychiatry 1997;63:611-5. [Crossref] [PubMed]
73. Räty S, Ruuth R, Silvennoinen K, Sabel BA, Tatlisumak T, Vanni S. Resting-state Functional Connectivity After Occipital Stroke. Neurorehabil Neural Repair 2022;36:151-63. [Crossref] [PubMed]
74. Kim YH, Cho AH, Kim D, Kim SM, Lim HT, Kwon SU, Kim JS, Kang DW. Early Functional Connectivity Predicts Recovery from Visual Field Defects after Stroke. J Stroke 2019;21:207-16. [Crossref] [PubMed]
75. Wang H, Huang Y, Li M, Yang H, An J, Leng X, Xu D, Qiu S. Regional brain dysfunction in insomnia after ischemic stroke: A resting-state fMRI study. Front Neurol 2022;13:1025174. [Crossref] [PubMed]
76. Serban CA, Barborica A, Roceanu AM, Mindruta I, Ciurea J, Pâslaru AC, Zăgrean AM, Zăgrean L, Moldovan M. A method to assess the default EEG macrostate and its reactivity to stimulation. Clin Neurophysiol 2022;134:50-64. [Crossref] [PubMed]
77. Vicentini JE, Weiler M, Casseb RF, Almeida SR, Valler L, de Campos BM, Li LM. Subacute functional connectivity correlates with cognitive recovery six months after stroke. Neuroimage Clin 2021;29:102538. [Crossref] [PubMed]
78. Park JY, Kim YH, Chang WH, Park CH, Shin YI, Kim ST, Pascual-Leone A. Significance of longitudinal changes in the default-mode network for cognitive recovery after stroke. Eur J Neurosci 2014;40:2715-22. [Crossref] [PubMed]
79. Bonkhoff AK, Schirmer MD, Bretzner M, Etherton M, Donahue K, Tuozzo C, Nardin M, Giese AK, Wu O, D, Calhoun V, Grefkes C, Rost NS. Abnormal dynamic functional connectivity is linked to recovery after acute ischemic stroke. Hum Brain Mapp 2021;42:2278-91. [Crossref] [PubMed]
80. Obayashi S. Cognitive and linguistic dysfunction after thalamic stroke and recovery process: possible mechanism. AIMS Neurosci 2022;9:1-11. [Crossref] [PubMed]
81. Newhart M, Trupe LA, Gomez Y, Cloutman L, Molitoris JJ, Davis C, Leigh R, Gottesman RF, Race D, Hillis AE. Asyntactic comprehension, working memory, and acute ischemia in Broca's area versus angular gyrus. Cortex 2012;48:1288-97. [Crossref] [PubMed]
82. Zhao L, Biesbroek JM, Shi L, Liu W, Kuijf HJ, Chu WW, Abrigo JM, Lee RK, Leung TW, Lau AY, Biessels GJ, Mok V, Wong A. Strategic infarct location for post-stroke cognitive impairment: A multivariate lesion-symptom mapping study. J Cereb Blood Flow Metab 2018;38:1299-311. [Crossref] [PubMed]
83. Ilves N, Lõo S, Ilves N, Laugesaar R, Loorits D, Kool P, Talvik T, Ilves P. Ipsilesional volume loss of basal ganglia and thalamus is associated with poor hand function after ischemic perinatal stroke. BMC Neurol 2022;22:23. [Crossref] [PubMed]