Research on dynamic functional connectivity and graph theoretic features of different motor subtypes in Parkinson’s disease based on ICA

1 Introduction

Parkinson’s disease (PD) is the second most common neurodegenerative disease worldwide, following Alzheimer’s disease (AD). Biologically, its main pathological characteristics include progressive loss of substantia nigra-striatal neurons and intracellular Lewy body aggregation. Notably, clinical symptoms do not manifest until approximately 50–60% of corresponding neuronal cells have died (1). Some studies (2) indicate that PD can typically be diagnosed when 60% of substantia nigra-striatal neurons and 80% of dopaminergic neurons in the striatum are lost, highlighting the urgent need for early diagnosis of PD. Resting-state functional magnetic resonance imaging (rs-fMRI) has become a commonly used tool for researching PD, as it can detect abnormal changes in brain functional activity and functional connectivity (FC) in PD patients (3, 4). This study focuses specifically on the postural instability/gait difficulty (PIGD) subtype and tremor-dominant (TD) subtype of PD, employing independent component analysis (ICA) (5)—a data-driven, statistical method that decomposes observed multidimensional signals into mutually statistically independent components by exploiting their independence and non-Gaussianity—as the core analytical method to decompose blood oxygen level-dependent (BOLD) signals from all brain voxels into spatially and temporally independent signals, directly measuring interactions within and between multiple brain networks, thereby grouping brain regions with higher correlation levels into functional brain network areas (6). On this basis, we combined dynamic functional network connectivity (dFNC) analysis methods based on ICA to calculate differences in dynamic connection strength over time between different independent brain networks and their temporal evolution features (7); additionally, we introduced dynamic graph theory methods to model brain networks using graph structures (with brain regions as nodes and neural connections as edges), calculating dynamic graph theoretic indicators such as characteristic path length, clustering coefficient, and node degree, further quantifying the dynamic topological structure characteristics of PIGD subtype, TD subtype, and healthy control (HC) group brain networks, and comprehensively elucidating the dynamic change patterns of brain functional networks in different PD subtypes.

PD patients with different motor subtypes exhibit significant heterogeneity in motor symptoms, cognitive function, and non-motor symptoms. Although their pathological basis centers on damage to the nigrostriatal pathway, motor control and cognitive regulation in the body are not accomplished by a single brain region alone; rather, they depend on the collaborative coordination of multiple brain networks, including the motor network, default mode network, and frontoparietal executive network (8). Although both subtypes share nigrostriatal pathway damage as their pathological basis, motor and cognitive control rely on the coordinated interaction of multiple brain networks. Notably, the two subtypes differ markedly in clinical symptom profiles—TD patients present predominantly with resting tremor, whereas PIGD patients exhibit more severe axial symptoms, bradykinesia, cognitive impairment and faster disease progression. These phenotypic differences suggest that the underlying neural circuit involvement—and hence the pattern of functional connectivity among the corresponding brain networks—may differ between the two subtypes. Accordingly, we hypothesized that TD and PIGD subtypes would exhibit subtype-specific alterations in functional connectivity and topological organization, which could be captured by dynamic connectivity and graph-theoretic analyses.

We chose ICA because it can decompose whole-brain BOLD signals into spatially independent components corresponding to distinct functional networks, allowing direct measurement of interactions within and between networks. ICA-based dFNC has been widely applied in resting-state fMRI studies and has successfully revealed time-varying connectivity states in both healthy populations and patients with neuropsychiatric and neurodegenerative disorders (9, 10), underscoring its utility for characterizing dynamic brain network alterations. In parallel, graph-theoretic analysis is an established framework for quantifying brain network topological properties across neurological diseases (11). Combining ICA-derived dFNC with dynamic graph metrics therefore offers a more comprehensive, complementary view of both temporal dynamics and topological structure than either static connectivity or conventional single-metric analyses alone.

Given this background, do PIGD subtype and TD subtype exhibit specific differences in connection patterns and topological structures of brain functional networks? Are these differences closely related to their specific clinical symptom profiles? This study aims to explore this in depth through neuroimaging approaches, providing theoretical foundations for precise subtyping and mechanistic research of different PD subtypes.

2 Methods

2.1 General information

Patients with PD diagnosed at Northern Jiangsu People’s Hospital from January 2021 to May 2024 were recruited, along with healthy subjects matched for age, sex, and education level. This study was approved by the Ethics Committee of Northern Jiangsu People’s Hospital and registered through the Chinese Clinical Trial Registry (registration number: ChiCTR2000041356). All subjects signed informed consent forms prior to enrollment.

2.2 Case selection

Inclusion criteria for PD patients: (1) Diagnosed with primary PD according to the Chinese Diagnostic Criteria for PD (2016 edition) (12); (2) All were right-handed.

Exclusion criteria: (1) Patients with Parkinson’s plus syndromes such as multiple system atrophy or progressive supranuclear palsy; (2) Patients with history of intracranial hemorrhage, cerebral infarction, head trauma, tumors, or neurosurgery; and (3) Patients with contraindications for MRI scanning. All PD patients underwent MRI scanning and clinical scale assessment after at least 12 h of discontinuation of anti-PD medications. Clinical assessment scales included: Unified Parkinson’s Disease Rating Scale Part III (UPDRS-III), Hoehn-Yahr staging, Mini-Mental State Examination (MMSE), and Montreal Cognitive Assessment (MoCA).

2.3 PD patient grouping criteria

According to the method proposed by Jankovic et al. (13) the mean TD score and the mean PIGD score were calculated separately using the UPDRS (mean tremor score: items 16 of Part II and items 20 and 21 of Part III; mean PIGD score: items 13–15 of Part II and items 29 and 30 of Part III), and the ratio between the two (TD/PIGD ratio) was used to classify motor subtypes. In the present study, this method was referenced and modified as follows: a single cutoff of 1.0 for the TD/PIGD ratio was used as the sole boundary (ratio ≤1.0 classified as the PIGD group, and ratio >1.0 classified as the TD group), without a separate indeterminate (mixed) group—that is, patients in the intermediate range (1.0–1.5) of the original classification were incorporated into the TD group. During the same period, volunteers without neurological or psychiatric diseases matched for age and sex were recruited from the community as the HC group. Ultimately, 36 PD patients were included: 19 patients with PIGD subtype, 17 patients with TD subtype, and 19 age-, sex-, and education-matched healthy subjects.

2.4 Functional magnetic resonance imaging data acquisition

Data was acquired using a Discovery MR750 3.0 T magnetic resonance scanner manufactured by GE Medical Systems, United States, equipped with a 32-channel head coil. Subjects were instructed to keep their eyes closed during scanning, remain awake and relaxed, and refrain from active thought. Foam pads were used to fix the head and minimize head movement. All patients underwent both 3D-T1 and resting-state BOLD sequence acquisition. Parameters were as follows:3D-T1 images were obtained using a magnetization preparation rapid gradient echo sequence: repetition time = 1900 ms, echo time = 2.13 ms, inversion time = 900 ms, slice thickness = 1 mm, flip angle = 9°, field of view = 256 mm × 256 mm, voxel size = 1 mm × 1 mm × 1 mm, total images = 176. BOLD sequences used echo-planar imaging for T2”weighted imaging: repetition time = 2000 ms, echo time = 30 ms, field of view = 225 mm × 225 mm, flip angle = 90°, matrix = 64 × 64, slice thickness = 3.5 mm, inter-slice gap = 0.7 mm, voxel size = 3.5 mm × 3.5 mm × 4.2 mm, number of slices = 31, total images = 179.

2.5 rs-fMRI data preprocessing

Resting-state functional magnetic resonance imaging (rs-fMRI) data were preprocessed using DPABI v4.1, a toolbox based on Statistical Parametric Mapping 12 (SPM12). The preprocessing pipeline was as follows: (1) Format conversion of raw imaging data; (2) The first 10 time points were discarded to minimize the effects of unstable magnetic field at the beginning of scanning and subjects’ adaptation to the scanning environment; (3) Slice-timing correction was performed to compensate for temporal differences in slice acquisition; (4) Realignment for head-motion correction. Subjects with translation >5 mm or rotation >5° in any direction were excluded to mitigate potential confounding effects of head motion; (5) Spatial normalization. The DARTEL algorithm was applied for spatial registration and normalization to warp functional images into standard MNI space; (6) Spatial smoothing using a Gaussian kernel with a full width at half maximum (FWHM) of 6 mm; (7) Removal of linear drift to reduce non-neuronal components such as time-dependent low-frequency drift during scanning; (8) Nuisance covariates including white matter signal, cerebrospinal fluid signal, and Friston-24 head-motion parameters were regressed out by linear regression. Global signal was not regressed in the present study to avoid spurious alterations in inter-regional functional correlations introduced by global signal regression; (9) Band-pass filtering within 0.01–0.08 Hz was carried out to suppress low-frequency drift and high-frequency physiological noise, while preserving the dominant low-frequency fluctuations of resting-state BOLD signals; and (10) Data quality control, including visual inspection of image registration and spatial normalization, assessment of head-motion parameters, and exclusion of subjects who failed the predefined head-motion criteria.

2.6 Independent component analysis

Preprocessed rs-fMRI data were decomposed using GIFT v4.0b software.¹ For ICA decomposition, the BOLD data from all participants (36 PD patients and 19 healthy controls, N = 55) were concatenated and jointly decomposed into 30 independent components, ensuring that the resulting components were derived from the same subject space and that the identified networks were comparable across the three groups. ICA decomposes data into spatially independent components (ICs), each with unique spatial maps and temporal profiles. The specific process was as follows: first, the minimum description length (MDL) criterion embedded in GIFT was used to estimate the dimensionality of independent components, and the final number of decomposed independent components was set to 30. MDL is an information-theoretic criterion that balances model goodness-of-fit against model complexity, selecting the number of components that minimizes the total description length of the data, thereby avoiding both over- and under-fitting (14). Next, the Infomax algorithm was used to decompose group-level fMRI data into 30 ICs. The ICASSO algorithm² was employed to improve the stability of ICA decomposition by repeating ICA computation 100 times, and stable ICs were obtained based on clustering outputs. Subsequently, the 30 derived ICs were screened. Visual inspection was performed according to the spatial patterns of each IC. Artifactual ICs predominantly distributed in non-gray-matter regions including ventricles, white matter, blood vessels and brain edges, as well as components contaminated by obvious head-motion or other physiological noise, were excluded. For the remaining ICs, multiple regression analysis was implemented in GIFT between spatial maps of each IC and predefined large-scale functional brain network templates. Spatial regression coefficients were calculated to evaluate spatial matching between each IC and canonical brain networks. Manual identification and classification were performed by comprehensively considering regression coefficients and spatial characteristics of ICs. Multiple ICs with distinct spatial features corresponding to the same canonical functional network were all retained. After screening, 12 valid ICs were finally obtained, which were assigned to nine functional brain networks: Auditory Network (AN), Primary Visual Network (PVN), Higher Visual Network (HVN), Default Mode Network (DMN), Left Frontoparietal Network (LFPN), Right Frontoparietal Network (RFPN), Dorsal Attention Network (DAN), Ventral Attention Network (VAN), and Sensorimotor Network (SMN) (Figure 1). Subsequent dFNC analysis was performed using the time-courses of all 12 screened valid ICs. Time-varying correlations between time-courses of different ICs were computed to characterize dynamic FC patterns among functional network components.

2.7 Dynamic functional connectivity

dFNC for each subject was calculated using the sliding window method and k-means clustering method. A sliding time window of 30 TR was applied to segment resting-state time series data. The previously identified 9 network components were used as regions of interest (ROI), with step size = 1 TR, to calculate dFNC for each subject. The resulting dFNC matrices were clustered using K-means clustering method, with cityblock (Manhattan distance) as the clustering distance metric. According to the elbow criterion, optimal clustering resulted in 4 classes of states, with 1,000 iterations. All matrices for each subject were assigned to one of the 4 specific states.

Time state indicators for each subject were calculated to assess temporal attributes of dFNC states, including: (1) State Frequency (F): ratio of time windows where a specific state occurred to total time windows; (2) Number of Transitions (NT): total number of transitions between different states; (3) Mean Duration Time (MDT): number of time windows a state appeared divided by total occurrences of that state; and (4) Transition Matrix (TM): state transition probability matrix.

2.8 Dynamic graph theory

After rs-fMRI data preprocessing, the Automated Anatomical Labeling-116 (AAL-116) atlas was used for whole-brain parcellation, yielding a total of 116 regions of interest (ROIs). The mean BOLD time-series across all voxels within each ROI were extracted as the representative signal for that brain region. The sliding-window approach was then applied to construct dynamic FC networks, with a window length of 30 TR and a step size of 1 TR. Within each sliding window, pairwise Pearson correlation analyses were performed on the mean BOLD time-series of the 116 ROIs to generate a symmetric 116 × 116 FC matrix. For subsequent graph-theoretical analyses, only positive correlations were retained. Negative connections with Pearson correlation coefficients below zero were set to zero. GRETNA (Graph Theoretical Network Analysis Toolbox) was utilized to construct undirected binary networks based on positive correlations. To mitigate the influence of arbitrary single-threshold selection, the network sparsity range was set from 0.05 to 0.40 with an interval of 0.01. The FC matrix was binarized at each sparsity threshold: retained connections were assigned a value of 1 and remaining connections were set to 0. Global and nodal graph-theoretical metrics were computed correspondingly under each sparsity level. Furthermore, GRETNA was used to calculate the area under the curve (AUC) for each graph-theoretical metric across the sparsity range of 0.05–0.40, producing integrated topological metrics relatively independent of specific threshold choices. For each subject, graph-theoretical metrics were obtained for every sliding window. The temporal variability of each metric across windows was further calculated to characterize dynamic fluctuations in brain network topological properties. These derived metrics were submitted to subsequent statistical analyses. The k-means clustering algorithm was implemented using the cityblock (L1) distance metric rather than Euclidean distance. The cityblock metric calculates the absolute difference between feature vectors and is less sensitive to outlier values, rendering it suitable for high-dimensional neuroimaging data (9, 15). To determine the optimal cluster number k, we adopted the elbow criterion based on the ratio of dispersion (within-group / between-group), serving as our cluster validity index. This index quantifies the balance between within-cluster compactness and between-cluster separation. The elbow criterion identifies the optimal k by locating the point where the rate of reduction in the dispersion ratio slows markedly, representing a trade-off between model complexity and clustering quality. We evaluated k values ranging from 2 to 10. The corresponding elbow plot illustrating the ratio of dispersion across tested k values was provided as Supplementary Figure S1, which identified k = 4 as the optimal cluster number.

2.9 Statistical analysis

Clinical data were analyzed using SPSS version 31.0 software.³ Continuous data approximately following normal distribution were expressed as mean ± standard deviation (x̄± s), with two-sample t-test used for comparison between 2 groups and one-way analysis of variance (ANOVA) for comparisons among 3 groups or more, followed by LSD-t test for further pairwise comparisons. Non-normally distributed continuous data were expressed as median [first quartile, third quartile] ([M(Q1, Q3)]), with Mann–Whitney U test for 2-group comparisons and Kruskal-Wallis H test for multi-group comparisons, followed by Bonferroni correction for pairwise comparisons. Categorical data were expressed as number (%). The Chi-square (χ²) test used for inter-group comparisons. p < 0.05 was considered statistically significant. Group analysis of dFNC was performed using the Stats module in GIFT software package. Inter-group comparisons of temporal attributes and connection strength under different states were performed using the general linear model (GLM), with age, sex, years of education included as covariates. Multiple comparison correction was performed using the false discovery rate (FDR). Dynamic graph theory analysis was based on 116 ROIs delineated by the AAL116 brain region template. Pearson correlation coefficients between 116 ROI time series were calculated pairwise within each time window to construct a 116 × 116 FC matrix. Brain network graph theory attributes (including global metrics and node metrics) were calculated using GRETNA based on this matrix. To reduce the impact of network density selection on results, network construction was performed using sparsity threshold ranges from 0.05–0.40 (step size 0.01), and area under the curve (AUC) of graph theory indicators within this threshold range was calculated. Variability of graph theory indicators across all time windows for each subject was further calculated.

GRETNA was used to calculate PD-brain network graph theory attributes, including global metrics and node metrics. Age, sex, and years of education were used as covariates, and partial correlation analysis was conducted to examine correlations between dFNC attributes (functional connection strength, time state indicators) of PIGD group patients and clinical scales. p < 0.05 was considered statistically significant.

3 Results

3.1 Comparison of clinical characteristics

A total of 55 participants were enrolled in this study, including 19 patients in the PIGD group, 17 patients in the TD group, and 19 individuals in the HC group. There were no significant differences in age, gender or education level among the three groups (p > 0.05). Statistically significant differences were observed in scores of the MMSE and MoCA (p < 0.05). All resting-state functional magnetic resonance imaging (rs-fMRI) scans were acquired under a unified off-medication state. Specifically, to minimize the potential confounding effect of dopaminergic medication (e.g., levodopa) on resting-state neural activity, all patients were asked to withhold antiparkinsonian medication for approximately 12 h (i.e., overnight withdrawal) before the fMRI session, and the time interval between the last dose and the scan was kept fixed at about 12 h across all participants. This design is in line with the widely used definition of the “practical OFF state” in PD neuroimaging studies, which commonly corresponds to dopaminergic withdrawal of ≥12 h (overnight). In addition, the levodopa equivalent daily dose (LEDD) of each patient was recorded to characterize the daily dopaminergic medication load.

No significant difference in disease duration was found between the PIGD and TD groups. However, significant intergroup differences were detected in the Hoehn-Yahr stage and motor subscore of the UPDRS-III (all p < 0.05), with higher scores in the PIGD group than in the TD group. Post-hoc pairwise comparisons of cognitive scales revealed that the MMSE scores decreased sequentially from the HC group to the TD group and then to the PIGD group, and the PIGD group had markedly lower MMSE scores than the HC group, with all differences reaching statistical significance. For MoCA scores, significant differences were only found between the two PD subgroups and the HC group, whereas no significant difference existed between the PIGD and TD groups. Detailed data are presented in Table 1.

Table 1. Demographic and clinical characteristics of PIGD group, TD group and HC group.

MMSE, Mini-Mental State Examination; UPDRS-III, Part III of the Unified PD Rating Scale; LEDD, Equivalent daily dose of levodopa.

3.2 Between-group component difference analysis within brain networks

ANOVA was performed based on the GLM using SPM12 software to compare the differences in spatial distribution of component maps across the 9 brain networks among the HC group, PIGD subtype group, and TD subtype group, with age and sex included as covariates to eliminate their confounding effects. The statistical thresholds were set at voxel level p < 0.01 and cluster level p < 0.05, corrected by Gaussian Random Field (GRF) theory. Post-hoc comparisons among the three groups were conducted using Tukey–Kramer test (Figure 2), as detailed in Table 2. All pairwise contrasts were pre-specified: HC vs. PIGD, HC vs. TD, and TD vs. PIGD. Only HC vs. PD-subtype contrasts reached statistical significance for the right middle temporal gyrus and right angular gyrus components; TD versus PIGD was not statistically significant for the right angular gyrus.

Table 2. Information of brain regions with significant differences in functional network components among groups.

3.3 Functional connectivity differences between brain networks

One-way analysis of variance was performed on the FC matrices among the 9 brain networks, with FDR correction applied to the results (p < 0.05). The ANOVA results revealed a significant between-group difference between the AN and SMN (F = 7.548, FDR-corrected p = 0.00137) (Figure 3A). Post-hoc tests demonstrated that FC was strongest in the HC group, followed by the TD group, and weakest in the PIGD group (Figure 3B).

3.4 Clustering analysis and dFNC state results

K-means clustering was employed, and four recurring functional connectivity (FC) states were ultimately identified. Across all sliding windows of the full cohort, the overall occurrence frequencies ranked from highest to lowest were: State 2 (38%), State 3 (23%), State 4 (21%), and State 1 (18%). Among the four states, both State 2 and State 3 exhibited a low-connectivity network pattern, with particularly pronounced reductions in FC between the SMN and the DMN, LFPN, and RFPN. In State 1, the DMN demonstrated high-intensity positive FC with bilateral frontoparietal networks (LFPN/RFPN) and the VAN. State 4 showed synchronized increases in FC among the VAN, SMN, and PVN, with enhanced network collaborative activation (Figure 4).

3.5 Comparison of dFNC state patterns across brain networks

Comparison of temporal attributes of dFNC states among the three groups showed that state occurrence frequency differed significantly only in State 4 among the three groups (F = 4.855, FDR-corrected p = 0.047). Post-hoc tests indicated that the occurrence frequency of State 4 was significantly higher in the HC group compared to the PIGD group, and exhibited a trend toward being higher than that in the TD group, whereas no significant difference was observed between the two PD subtypes. State transition count: No significant difference in state transition counts was found across the three groups (F = 0.515, p = 0.600), suggesting no significant difference in overall dynamicity of brain network state transitions among the three groups. Mean dwell time: Differences in mean dwell time for State 4 were observed among the three groups at uncorrected level only (F = 3.376, p = 0.042), but did not reach statistical significance after FDR correction. Post-hoc tests revealed an overall trend with the HC group showing higher values than both PD subtypes, but no significant differences between any pair of groups. State transition probability matrix: Transition probabilities among the four states showed no significant differences among the three groups (non-significant before and after correction) (Figure 5).

3.6 Analysis of topological properties of brain functional networks

3.6.1 Global metrics analysis

ANOVA was performed on functional connections across all brain nodes. At a significance threshold corrected by FDR (p < 0.05), no brain regions showed significant between-group differences. The comparisons of global graph-theoretical network attributes for the three groups are detailed in Table 3.

Table 3. Comparison of global attributes of the three groups.

Eg, global efficiency; Eloc, local efficiency; Cp, clustering coefficient; Lp, characteristic path length.λ, normalized characteristic path length; γ, normalized clustering coefficient; σ, small-worldness; In this study, the global topological properties refer to the AUC of the metrics calculated across different sparsity thresholds (0.05–0.40). The AUC was used to integrate results from multiple thresholds and avoid bias caused by selecting a single threshold.

3.6.2 Node metric analysis

One-way ANOVA was performed on the graph theoretical properties of node maps across the three groups. No significant differences were found after FDR correction; however, at the uncorrected p < 0.05 threshold, differences among the three groups exhibited the following regional distribution patterns: Cuneus_R demonstrated differential trends in both Betweenness Centrality (BC) and Node Efficiency (NE), while Precuneus_R showed differential trends in both Degree Centrality (DC) and NE, as detailed in Tables 4, 5 and Figure 6. Further Tukey–Kramer post-hoc tests revealed that the NE of Cuneus_R in the PIGD group was significantly higher than that in the HC group (uncorrected p < 0.05); additionally, the DC of Precentral_R in both the PIGD group and TD group were significantly higher than in the HC group (respectively uncorrected p < 0.05 and uncorrected p < 0.01). The brain regions exhibiting differential trends in BC were primarily distributed in the frontal lobe, occipital lobe, temporal lobe, and cerebellum; the regions showing differential trends in DC were concentrated in the frontal lobe, parietal lobe, temporal lobe, and cerebellum, particularly involving multiple subregions of the cerebellar hemispheres and vermis; NE differential trend regions were mainly located in the occipital lobe and parietal lobe, with partial involvement of temporal lobe regions (Figure 7).

Table 4. Summary of intergroup differences in nodal topological properties.

PVN, Primary Visual Network; HVN, High-order Visual Network; DAN, Dorsal Attention Network; AN, Auditory Network; LFPN, Left Fronto-Parietal Network; SMN, Sensorimotor Network; DMN, Default Mode Network; CN, Cerebellar Network.

Table 5. Comparison of node attributes of the three groups.

BC, Betweenness Centrality; DC, Degree Centrality; NCp, Node Clustering Coefficient; NLe, Node Local Efficiency; Ne, Node Efficiency.

3.7 Correlation analysis between dFNC metrics and clinical characteristics

After adjusting for the effects of age, sex, and years of education, the fractional occupancy of State 2 in the TD group showed a positive correlation with Hoehn-Yahr staging scores (r = 0.506, p = 0.038). In the PIGD group, the mean dwell time in State 3 demonstrated a negative correlation with UPDRS-III motor scores (r = −0.465, p = 0.045). No significant correlations were observed between other temporal property metrics, FC strength metrics, and clinical characteristic scores (p > 0.05) (Figure 8).

4 Discussion

4.1 Differences between PD group and HC group

In the present study, ICA combined with dFNC and dynamic graph theory was adopted to compare the dynamic characteristics of brain functional networks among the PIGD subtype, TD subtype and HC group from a dynamic perspective. The results revealed that compared with the HC group, the overall PD group (PIGD and TD subgroups) exhibited significant abnormalities in the spatial distribution of five core brain networks, namely the AN, LFPN, RFPN, VAN and SMN. Specifically, component values of the right middle temporal gyrus within the AN and VAN, as well as the right middle cingulate gyrus and right supplementary motor area within the SMN, were markedly higher in PD subtypes (TD and PIGD) than those in the HC group. These findings may reflect compensatory overactivation of brain networks subserving auditory information processing, attentional function and sensorimotor regulation in PD patients, consistent with core clinical manifestations of PD including hypophonia, slowed auditory response, bradykinesia and muscular rigidity. The results were also in line with previous studies (16, 17), confirming that abnormal spatial distribution of brain networks serves as a crucial neuroimaging basis for functional deficits in PD. In terms of FC, the strength of FC between AN and SMN was significantly reduced in the PD group relative to the HC group, suggesting impaired coordination of auditory-motor integration. This may account for speech-motor disorders and desynchronization between movements and auditory signals in PD patients. dFNC analysis demonstrated that the overall PD group showed a notably lower occurrence rate of State 4 (characterized by enhanced synergistic activation of VAN, SMN and PVN), while States 2 and 3 (low-connectivity network patterns) dominated the functional states. These observations implied a decline in the synergistic activation capacity of brain networks in PD patients, with low synergistic activation patterns becoming the predominant brain functional state, which may underlie the impaired coordination of cognitive and motor functions in PD. Multiple prior studies have verified that PD patients spend substantially more time in sparse/low-connectivity states (State 2 and 3) and exhibit lower occurrence rates of highly connected states (State 1 and 4) compared with healthy individuals (18, 19). For nodal graph theoretical metrics, no significant intergroup differences were observed across the three groups after FDR correction. However, trends of differences in topological properties of brain regions were detected at the uncorrected statistical level. The nodal efficiency (NE) of the right cuneus (Cuneus_R) was significantly higher in the PIGD group than in the HC group. Meanwhile, the degree centrality (DC) of the Right Precuneus was elevated in both PIGD and TD groups relative to the HC group. These results suggested abnormal activation of the motor cortex in PD patients, which may be associated with motor dysfunction (20). In summary, the overall differences in brain function between the HC group and PD group were characterized by abnormal spatial distribution of multiple brain networks, weakened FC of specific networks, declined synergistic activation capacity of dFNC, and elevated activation of the motor cortex. Collectively, these brain functional abnormalities constitute the neuroimaging substrate for cognitive and motor impairments in PD, providing objective evidence for the general diagnosis of this disease.

4.2 Differences between two motor subtypes of PD and their correlations with cognitive and motor scale scores

PD presents prominent clinical heterogeneity and is clinically classified into two motor subtypes based on predominant motor symptoms: TD and PIGD. Previous studies have demonstrated that the PIGD subtype is mainly manifested as postural and gait disturbances as well as bradykinesia, whereas the TD subtype is dominated by resting tremor. With regard to cognitive function, the PIGD subtype suffers from more severe executive dysfunction, while cognitive impairment is relatively mild in the TD subtype (21). Using rs-fMRI (rs-fMRI), the present study identified subtype-specific alterations in brain network features, which were strongly correlated with clinical scale scores. These findings uncovered the distinct neuropathological mechanisms underlying varied clinical phenotypes of the two subtypes.

4.2.1 Subtype-specific differences in spatial distribution of brain networks

Our results showed that the component value of the left superior frontal gyrus within LFPN was significantly lower in the PIGD group than in the TD group, which was consistent with previous reports. As a core network for executive control, LFPN is involved in higher-order cognitive functions such as working memory, inhibitory control and task switching. Accumulated evidence has indicated that LFPN dysfunction correlates with the severity of cognitive impairment, particularly executive dysfunction, in PD patients (22). In this study, the PIGD group obtained lower MMSE scores than the TD group, indicating more severe cognitive deficits, which corroborated the attenuated LFPN function in the PIGD subgroup. Nevertheless, no significant between-group difference was found in MoCA scores. This phenomenon may be explained by two factors. First, the limited sample size (19 PIGD patients and 17 TD patients) may have insufficient statistical power to detect subtle cognitive differences between the two subtypes. Second, although MoCA is sensitive to executive dysfunction, it may fail to capture mild cognitive changes induced by specific LFPN lesions. Further studies combining specialized executive function assessments such as the Frontal Assessment Battery (FAB) are required to validate the quantitative association between frontoparietal network function and cognitive impairment. The component value of the right middle frontal gyrus in RFPN exhibited a gradual decreasing trend across groups: HC group > TD group > PIGD group. This gradient pattern was highly consistent with higher Hoehn-Yahr staging and Unified PD Rating Scale Part III (UPDRS-III) motor scores in the PIGD group. RFPN participates in motor planning, sequence programming and movement execution, and its dysfunction directly leads to postural gait disorders and bradykinesia. Systematic neuroanatomical reviews have documented subtype-specific alterations in white matter microstructure in PIGD patients, including reduced nerve fiber density in the corona radiata and multiple white matter tracts, which are more extensive compared with the TD subtype (23, 24). Given that postural instability and gait difficulty are the hallmark symptoms of the PIGD subtype, the markedly reduced RFPN function in this subgroup suggests that RFPN may serve as a relevant neuroimaging indicator for differentiating two PD subtypes. This finding also provides a neurophysiological explanation for the clinical features of PIGD, including faster disease progression and higher fall risk. The component value of the right angular gyrus within the VAN showed a gradient pattern of HC > PIGD > TD. Significant differences were detected only between HC and each PD subtype, while no statistically significant difference was observed between TD and PIGD. This reduced component value in TD patients may imply that tremor occurrence may be linked to attenuated VAN activation. As a key component of the attentional control network, VAN is responsible for attentional shifting, multimodal information integration and sensorimotor coordination. Existing studies have reported abnormal activation of the basal ganglia-thalamocortical circuit involving the sensorimotor cortex and attention networks in PD patients with tremor (25). The abnormal activation of the right angular gyrus in VAN observed in the present study may be associated with dysregulated automatic motor control and abnormal allocation of attentional resources related to tremor, offering new theoretical insights into the neural mechanisms of tremor.

4.2.2 Subtype characteristics at the functional connectivity

Level Further analysis revealed that the FC strength between AN and SMN followed the trend of HC group >TD group >PIGD group, with the lowest connectivity detected in the PIGD group. This result was in accordance with more severe bradykinesia in PIGD patients. The AN-SMN FC mediates the integration of auditory input and motor output, and plays an essential role in speech-motor control, movement sequence execution and posture regulation. Prior studies have demonstrated that PIGD patients present a specific dynamic functional network pattern characterized by decoupling among the default mode network, AN and SMN. Moreover, the strength of AN-SMN connectivity is positively correlated with motor function scores (26, 27). The pronounced reduction of AN-SMN connectivity in the PIGD group may reflect more extensive impairment of neural integration, resulting in declined accuracy of postural gait control and speech-motor performance. Consistent with previous literature, this finding suggests that AN-SMN FC can be used to evaluate the severity of motor dysfunction in PIGD patients.

4.2.3 Dynamic functional connectivity patterns

dFNC analysis showed that both PD subtypes had a lower occurrence rate of State 4 relative to the HC group, indicating impaired brain functional plasticity and synergistic activation capacity in all PD patients. This result aligns with the theory of compromised neural plasticity during PD progression. Previous research has confirmed that the temporal properties of dFNC states differ significantly between PD patients and HC. The number of brain network state transitions increases alongside cognitive deterioration, reflecting progressive decline of synergistic activation (28). Notably, no significant difference in this metric was observed between the two subtypes, indicating that reduced synergistic activation is a universal feature of PD rather than a subtype-specific alteration. This finding is consistent with the clinical manifestation that both subtypes suffer from functional deficits but differ in the types of impaired functions.

This study explored potential neuroimaging features derived from ICA-based dynamic functional connectivity and graph theory analysis to preliminarily characterize brain network alterations in PD patients and motor subtypes. Notably, most subtype-level differences observed in the present study were identified at the uncorrected statistical level and did not survive strict FDR correction. In addition, no significant pairwise differences were detected between the TD and PIGD subtypes for most network components and dynamic metrics. Therefore, all findings described in this section should be regarded as exploratory rather than validated diagnostic biomarkers. No imaging indicator identified in this work can currently support reliable clinical differentiation or definite subtype classification.

At the whole-group level, the decreased functional connectivity between the auditory network and sensorimotor network exhibited robust differences between PD patients and healthy controls after FDR correction, which may serve as a preliminary candidate feature reflecting impaired auditory-motor integration in PD. Similarly, the reduced occurrence rate of State 4 in dFNC analysis may tentatively reflect disrupted synergistic activation of multi-network coordination in PD, which could potentially indicate altered brain dynamic plasticity related to disease status.

At the subtype level, although several regional component values in the LFPN and RFPN showed trending differences between TD and PIGD patients, these discrepancies were not statistically rigorous after multiple comparison correction and lacked reliable pairwise subtype validation. Accordingly, these network alterations cannot be considered subtype-specific biomarkers. No stable tremor-related imaging feature was identified for TD subtype in the current cohort.

Collectively, the present neuroimaging results provide preliminary exploratory evidence for dynamic brain network abnormalities in PD and subtle subtype-associated functional alterations. However, given the limited sample size, cross-sectional design, uncorrected trending differences, and insufficient pairwise subtype contrasts, these features cannot yet be applied for clinical diagnosis or subtyping. Future large-sample, longitudinal studies are required to verify the stability, specificity, and translational value of these potential imaging indicators before clinical application.

5 Conclusion

In this study, we observed that PD patients (both TD and PIGD groups) exhibited spatial distribution abnormalities in multiple brain networks, including elevated component values of the right middle temporal gyrus in the AN and VAN, and of the right middle cingulate gyrus and right supplementary motor area in the SMN, together with decreased FC between the AN and SMN and a lower occurrence rate of State 4 in dFNC, relative to the HC group. The most consistent and robust finding across the study was therefore the divergence between PD patients and healthy individuals, supporting the view that a combination of ICA-derived dFNC and dynamic graph theory analysis can provide objective neuroimaging evidence for distinguishing PD patients from HC. With respect to the two motor subtypes (TD vs. PIGD), the between-group differences observed were more limited and less robust. We detected subtype-related differences in the component values of the left superior frontal gyrus within the LFPN and the right middle frontal gyrus within the RFPN. However, these subtype differences were not observed consistently across all analytical levels; for example, no significant difference was found between the two subtypes in the occurrence rate of State 4 in dFNC. Owing to the modest sample size and the cross-sectional design of this study, the present evidence is insufficient to conclusively establish subtype-specific neuroimaging markers. Accordingly, the capacity of these indicators to reliably differentiate the two motor subtypes should be regarded as preliminary, pending validation in larger, longitudinal cohorts.

In summary, the most solid conclusion drawn from the present study is the difference in dynamic brain network characteristics between PD patients and HC. The subtype-related alterations reported here should be interpreted with caution and warrant further investigation.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fneur.2026.1943343/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Jan Kassubek, University of Ulm, Germany

Reviewed by: Farzin Hajebrahimi, New Jersey Institute of Technology, United States

Renxi Li, George Washington University, United States

Footnotes

http://matlab.mrn.org/software/gift/

http://research.ics.tkk.fi/ica/icasso/

http://www.ibm.com/spss