Regional-specific structural and functional changes of posterior cerebellar vermis across different stages of Parkinson’s disease with gait dysfunction

Abstract

Stage-specific roles of posterior cerebellar vermis (PV) subdivisions, the posterior superior vermis (PSV) and posterior inferior vermis (PIV), in Parkinson’s disease postural instability/gait difficulty (PD-PIGD) remain unclear. This retrospective, cross-sectional study investigated their volumetric and functional connectivity (FC) changes and clinical correlates across PD-PIGD stages. We analyzed 94 PD-PIGD patients (Hoehn & Yahr, HY1-4) and 46 healthy controls (HCs). Patient data were from outpatients and baseline assessments in two clinical trials (ClinicalTrials.gov: NCT02969941, reg. 2016-06-01; NCT05192759, reg. 2021-11-22). Compared with HCs, early-stage (HY1) patients showed enhanced PSV-left paracentral lobule (L_PCL) FC, alongside a trend toward increased PSV volume. This PSV-L_PCL FC correlated with better cognition function and gait performance, an association partly cognition-mediated. Our findings reveal a PSV-specific nonlinear pattern of structural and functional changes in PD-PIGD, distinct from PIV or other cerebellar subregions, potentially reflecting early compensatory mechanisms transitioning to later network dysfunction.

Introduction

Gait impairments, encompassing postural instability and gait difficulty (PIGD), are cardinal and disabling features of Parkinson’s disease (PD) that profoundly diminish functional independence and quality of life¹. Compared with tremor-dominant (TD) phenotypes, the PIGD subtype poses a significant clinical challenge and is often associated with more rapid disease progression, reduced responsiveness to dopaminergic therapy, and a heightened risk of cognitive decline^2–4. The limited efficacy of standard treatments for PD-PIGD underscores the involvement of nondopaminergic systems, and a deeper understanding of the underlying neural mechanisms is needed to develop effective, targeted interventions⁵.

The understanding of PD is evolving from a primary basal ganglia disorder to a model of progressive multisystem neurodegeneration affecting distributed brain networks^6,7. Considering the cerebellum’s central functions in gait, balance, and increasingly recognized cognitive processes, understanding its specific involvement is key within the network model of PD-PIGD^2,6. Emerging evidence supports a ‘network-selective vulnerability’ framework in neurodegeneration, suggesting that pathology preferentially targets functionally interconnected brain regions^8,9. Cerebellar involvement in Alzheimer’s disease (affecting default mode network-connected regions such as Crus I/II) and frontotemporal dementia (affecting salience network-connected regions such as Lobule VI) exemplifies this principle⁸. Applying this principle to PD, which also involves progressive network dysfunction, requires an understanding of how specific cerebellar nodes that contribute to PIGD symptoms are affected throughout the disease course¹⁰.

The posterior cerebellar vermis (PV), which is integral to gait and postural control, is divisible into functionally distinct subdivisions: the posterior superior vermis (PSV), which comprises vermis VI, Crus I/II, and VIIb, and the posterior inferior vermis (PIV), which comprises vermis VIIIa/b, IX, and X^11,12. These subdivisions have divergent connectivity profiles that are critical to understanding their potential roles in PD-PIGD. The PSV is preferentially linked to cortical motor planning and cognitive control circuits, which are distinct from the predominant connections of the PIV to spinocerebellar and the vestibulocerebellar pathways that mediate sensorimotor integration and postural adjustments^13–17. This differential connectivity suggests distinct vulnerability patterns that warrant investigation in PD-PIGD.

The evidence suggests that cerebellar involvement in PD is dynamic, potentially reflecting nonlinear changes across disease stages rather than simple progressive atrophy. For example, functional connectivity (FC) of the PV with the sensorimotor cortex correlates with motor severity, and large-scale structural analyses suggest possible early cerebellar volume increases followed by later atrophy^18,19. Moreover, the interplay between cognitive and gait impairments is a prominent feature of PD-PIGD². Given the established connectivity of the PSV with cortical networks relevant to both cognition and motor control and the recognized dynamic nature of cerebellar involvement in PD, understanding the specific contribution of the PSV to these critical cognitive-motor interactions is therefore a key objective for elucidating PD-PIGD neural mechanisms.

Accordingly, this study aimed to characterize structural and functional alterations of the PSV and PIV across distinct PD-PIGD stages. We hypothesized that the PSV, given its connectivity profile with cortical circuits, would exhibit a nonlinear, stage-dependent pattern of change distinct from that of the PIV or other cerebellar subregions. Elucidating such differential vulnerability could establish PSV as a potential stage-dependent biomarker and a target for therapeutic interventions, such as noninvasive neuromodulation, aimed at improving both gait and cognition in PD-PIGD patients.

Results

Demographic and clinical characteristics

The final primary analysis included 94 PD-PIGD patients (mean age 62.4 ± 9.7 years, range 41–82; 47.9% female) and 46 HCs (mean age 61.1 ± 9.3 years, range 46–83; 58.7% female) (Fig. 1). The groups did not significantly differ in terms of sex distribution (p = 0.229) or years of education (p = 0.08). Compared with the participants in the HC group, the patients in the PD-PIGD group had significantly lower MoCA scores (p < 0.001) but did not differ significantly in age (p = 0.444). When the data were stratified by HY stage, significant differences in age were observed among the four groups (HC, HY1, HY2, and HY3-4) (F_(3,136) = 7.00, p < 0.001). In the PD-PIGD subgroups, increasing HY stage was associated with a significantly longer disease duration, higher LEDD, lower MoCA scores, lower MMSE scores, higher UPDRS-III scores, and higher PIGD scores (all p < 0.001). No significant differences across HY stages were found for HAMA or HAMD scores (all p > 0.5) (Table 1).

Fig. 1.

Flowchart depicting the participant inclusion process.

Table 1.

Demographic and clinical characteristics of the HCs and PD-PIGD patients

Group	HC (n = 46)	PD-PIGD (n = 94)	PD-PIGD (n = 94)			t Test	ANOVA
			HY1 (n = 33)	HY2 (n = 30)	HY3–4 (n = 31)	p value (HC vs. PD-PIGD)	p Value
Age, y	61.09 ± 9.26 (46–83)	62.40 ± 9.67 (41–82)	57.73 ± 9.09 (41–75)	61.97 ± 9.23 (42–78)	67.81 ± 8.05 (54–82)	0.444b	<0.001c
Sex (M/F)	19/27	49/45	20/13	14/16	15/16	0.229a	0.403a
Education, y	8.78 ± 4.28 (0–15)	7.33 ± 4.71 (0–17)	8.58 ± 5.19 (0–17)	6.73 ± 4.58 (0–16)	6.58 ± 4.14 (0–15)	0.08b	0.077c
MoCA	23.72 ± 3.09 (15–30)	20.18 ± 5.95 (10–30)	23.52 ± 4.41 (12–30)	20.93 ± 6.06 (10–29)	15.90 ± 4.66 (10–28)	<0.001b	<0.001c
Disease duration, y	–	–	2.27 ± 1.82 (0.5–5)	3.50 ± 2.77 (2–9)	5.18 ± 3.12 (3–11)	–	<0.001 d
LEDD, mg	–	–	244.92 ± 219.23 (0–580)	420.17 ± 140.35 (100-675)	540.08 ± 166.42 (200–1025)	–	<0.001 d
MMSE	–	–	27.70 ± 2.34 (21–30)	26.53 ± 3.91 (18–30)	23.77 ± 3.36 (16–29)	–	<0.001 d
HAMD	–	–	7.09 ± 4.70 (0–22)	7.57 ± 5.15 (0–23)	8.19 ± 5.33 (1–21)	–	0.684 d
HAMA	–	–	6.42 ± 3.35 (1–14)	7.47 ± 5.00 (0–25)	7.16 ± 4.22 (1–21)	–	0.599 d
UPDRS III score	–	–	20.82 ± 9.74 (8–39)	29.40 ± 11.42 (15–42)	37.23 ± 15.83 (20–82)	–	<0.001 d
PIGD score	–	–	3.94 ± 1.68 (2–7)	4.97 ± 2.21 (3–10)	9.87 ± 2.66 (7–16)	–	<0.001 d

^aChi-square test, ^bindependent samples t test between the HC and PD-PIGD groups, ^cANOVA of the HC, HY1, HY2, and HY3-4 groups,

^dANOVA of the HY1, HY2, and HY3-4 groups. Bold indicates statistical significance (p < 0.05). Continuous variables are presented as the means ± SD (minimum‒maximum), and categorical variables are presented as counts (n).

HAMA Hamilton Anxiety Scale, HAMD Hamilton Depression Scale, LEDD levodopa equivalent daily dose, MMSE Minimum Mental State Examination, MoCA Montreal Cognitive Assessment, PIGD postural instability and gait difficulty, UPDRS III Unified Parkinson’s Disease Rating Scale Part III.

Volumetric analyses

Comparisons between the overall PD-PIGD group and the HC group revealed no significant differences in volume for any analyzed cerebellar subregion (all p > 0.1). However, in the primary ANCOVA, adjusting for age, a significant effect of group was noted on PSV volume (F_(3,135) = 3.10, p = 0.029) and CBMc volume (F_(3,135) = 2.95, p = 0.035), but not on PV, PIV, or other cerebellar ROIs (all p > 0.1) (Fig. 2 and Supplementary Table 2). Post hoc comparisons revealed that the PSV volume in the HY1 group was significantly greater than that in the HC group (Cohen’s d = 0.66, p_bonf = 0.028) (Fig. 2f). Additionally, the CBMc volume was significantly smaller in the HY3-4 group than in the HY1 group (Cohen’s d = -0.80, p_bonf = 0.021). However, these specific volumetric findings were sensitive to further covariate adjustment. In the sensitivity analyses in which simultaneous adjustments were made for age, sex, and education years, neither the PSV difference between HY1 and HC (p = 0.096) nor the CBMc difference between HY3-4 and HY1 (p = 0.07) remained significant (Supplementary Table 3).

Fig. 2. Posterior vermis subregion diagrams and volumetric comparisons.

a PSV and PIV diagrams. Violin and box plots comparing b PV, c PSV, and d PIV volumes between the combined PD-PIGD and HC groups using independent samples t tests (t values and p values are shown). Violin and box plots comparing e PV, f PSV, and g PIV volumes across the four groups (HC, HY1, HY2, and HY3-4). The F values and p values shown for these four-group comparisons were obtained by one-way ANCOVA adjusted for age. Post hoc pairwise comparisons utilized Bonferroni correction. In panel f, an asterisk (*) indicates a significant difference (p < 0.05, Bonferroni-corrected, age-adjusted) for the comparison between the HY1 group and the HC group. Bold font for p values reported on the figure indicates statistical significance (p < 0.05). PIV posterior inferior vermis, PSV posterior superior vermis, PV posterior vermis.

FC analyses

An initial FC analysis comparing the overall PD-PIGD group with the HC group, using the entire PV as a seed, revealed a significant cluster of increased connectivity with the left paracentral lobule (L_PCL) (Supplementary Fig. S1). Subsequent analyses focusing on PV subdivisions revealed that this effect was primarily driven by the PSV. The four-group voxelwise GLM using the PSV seed identified significant overall group differences (F test map, Fig. 3b) in connectivity to five distinct clusters: the L_PCL (Somatomotor Network), right precentral gyrus (R_PCG; Somatomotor Network), right putamen (R_Put; Subcortical), right cerebellar Crus II (R_Crus II; Cerebellum Network), and left fusiform gyrus (L_FG; Limbic Network) (all cluster p_FDR < 0.05) (Supplementary Table 4). A pattern of decreasing FC between the PSV and L_PCL/R_PCG was observed across disease stages (Fig. 3c). Post hoc pairwise comparisons confirmed significantly greater PSV-L_PCL FC and PSV-R_PCG FC in the HY1 group than in the HC group (both p_FDR < 0.001). Compared with those in the HY1 group, the PSV-R_Put FC was significantly lower (p_FDR = 0.041), and the PSV-R_Crus II FC was significantly higher (p_FDR = 0.022) in the HY3-4 group. Additionally, the PSV-L_FG FC in the HY3-4 group was higher than that in the HC group (p_FDR = 0.042) (all pairwise comparisons detailed in Fig. 3d). These key findings regarding PSV FC alterations remained robust in the sensitivity analyses in which adjustments were also made for sex and education (Supplementary Tables 5 and 6). In contrast, no significant group differences were observed in the analysis originating from the PIV seed. Analyses seeded from the other cerebellar subregions (CBMm, Lobules I–IV, IX) revealed significant group differences emerging predominantly in the later (HY3–4) stage, thereby confirming the relative specificity of PSV alterations in early-stage (HY1) PD-PIGD (Supplementary Figs. S2–S4).

Fig. 3. PSV-seed FC differences between HCs and PD-PIGD patients across disease stages.

a Average seed-to-voxel FC maps for the PSV-seed within each group (HC, HY1, HY2, and HY3–4). Warm/cool colors indicate positive/negative correlations, thresholded for visualization. b Brain regions showing significant group differences in PSV-seeded FC, identified by a whole-brain voxelwise GLM comparing the four groups (F test map). Maps are thresholded at voxel-level p < 0.001 (uncorrected) and cluster-level p_FDR < 0.05. The color bar indicates the F-statistic values. c Bar graph showing the mean Fisher z-transformed FC values extracted from the significant clusters identified in (b) across the four groups (HC, HY1, HY2, and HY3–4); the error bars represent the standard error of the mean. d Post hoc pairwise group comparisons (HY1 vs. HC, HY3-4 vs. HC, and HY3–4 vs. HY1). The brain maps depict regions showing significant differences in connectivity (T-maps; warm colors: increased connectivity in the first group relative to the second; cool colors: decreased connectivity in the first group relative to the second). Statistical thresholds were set at a voxel-level of p < 0.001 (uncorrected) and cluster-level of p_FDR < 0.05. Cluster labels: a L_PCL (left paracentral lobule); b R_PCG (right precentral gyrus); c L_FG (left fusiform gyrus); d R_Put (right putamen); e R_Crus II (right cerebellar crus II).

Connectivity-guided therapeutic targeting of the PSV-L_PCL pathway

Motivated by the finding of altered PSV-L_PCL connectivity in our primary analysis and the spatial proximity of the identified L_PCL cluster (peak MNI: −8, −24, +74) to a previously described TMS target for PD-FoG (MNI: −10, −24, +75), we performed a retrospective validation analysis²⁰. Using baseline resting-state (rs) fMRI data and clinical outcome data from 24 PD-FoG participants who had received accelerated high-dose theta-burst stimulation (ah-TBS) targeting the L_PCL (see Table 2 for cohort characteristics), we calculated the Euclidean distance between the treated TMS target (hereafter, ‘actual target’) and an individualized ‘optimal target’ (defined as the voxel of peak baseline PSV-L_PCL FC within the L_PCL for each participant). We calculated the fractional improvement rate after ah-TBS for the 180° stand-and-spin (SS180) time, 5-m timed up-and-go (5mTUG) time, UPDRS III score, and PIGD score; this rate was determined for each measure by dividing the change in time or score (pre-intervention minus post-intervention) by the pre-intervention value. Using partial correlations controlling for age, we subsequently assessed the relationship between the z-standardized distance metric (concept illustrated in Fig. 4a) and the z-standardized fractional improvement rates.

Table 2.

Baseline demographic and clinical characteristics of the validation cohort

Group	PD-FoG (n = 24)
Age, y	69.17 ± 7.98 (51–82)
Sex, (M/F)	15/9
Education, y	7.00 ± 4.95 (0–16)
Disease duration, y	5.33 ± 2.96 (2–13)
HY	2.48 ± 0.63 (2–4)
LEDD, mg	559.94 ± 216.54 (0–950)
MMSE	25.50 ± 3.20 (18–29)
MoCA	17.83 ± 4.87 (12–28)
Distance, mm	9.63 ± 3.33 (4.36–16.55)
SS180 improvement rate	0.32 ± 0.30 (−0.25 to 0.89)
5mTUG improvement rate	0.28 ± 0.23 (−0.17 to 0.74)
UPDRS III improvement rate	0.18 ± 0.11 (0–0.36)
PIGD improvement rate	0.21 ± 0.27 (0–1.00)

Continuous variables are presented as the means ± SD (minimum‒maximum), and categorical variables are presented as counts (n). The data describe the 24 PD-FoG patients included in the retrospective validation analysis of TMS targeting the left paracentral lobule. Fractional improvement rates were calculated using the formula: (pre-intervention value minus post-intervention value) divided by the pre-intervention value. 5mTUG 5-m timed up-and-go time, LEDD levodopa equivalent daily dose, MMSE Minimum Mental State Examination, MoCA Montreal Cognitive Assessment, PD-FoG Parkinson’s disease with freezing of gait, PIGD postural instability and gait difficulty, SS180 180° stand-and-spin time, UPDRS III Unified Parkinson’s Disease Rating Scale Part III.

Fig. 4. Correlation between the target distance (z) and clinical improvement rate (z) in PD-FoG.

a Diagram illustrating the Euclidean distance measured between the actual TMS target (‘Actual’, MNI: −10, −24, +75) and the individualized optimal target based on peak PSV-L_PCL FC (‘Optimal’). b–e Scatter plots show partial correlations (controlling for age) between the standardized Euclidean distance (zDistance: representing the separation between the actual TMS target from the individual’s optimal PSV-L_PCL connectivity peak) and standardized fractional improvement rates (z) following 5 days of ah-TBS. Correlations are shown for: b 180° stand-and-spin time improvement rate (zSS180 rate). c 5-m timed up-and-go improvement rate (z5mTUG rate). d PIGD score improvement rate (zPIGD rate), and e UPDRS III score improvement rate (zUPDRS III rate). The partial correlation coefficient (r) and corresponding p value (adjusted for age) are displayed on each plot. A shorter distance was significantly correlated with greater improvements in the SS180 and 5mTUG tasks, controlling for age. Bold indicates statistical significance (p < 0.05) for displayed p-values.

A significant negative correlation was found between the z-standardized Euclidean distance and the z-standardized improvement rate for SS180 (r = −0.67, p < 0.001) and 5mTUG (r = −0.50, p = 0.013) (Fig. 4b, c). No significant correlation for improvement in the PIGD score (r = −0.36, p = 0.083) or the UPDRS III score (r = −0.35, p = 0.098) was observed (Fig. 4d, e). Sensitivity analyses, additionally controlling for sex and education years alongside age, confirmed the robustness of these significant findings (SS180: r = −0.67, p < 0.001; 5mTUG: r = −0.49, p = 0.026). These results suggest that stimulating closer to an individual’s peak PSV-L_PCL connectivity locus (optimal target) is associated with greater improvement in specific gait tasks (SS180 and 5mTUG).

Analyses of the correlations between PSV metrics and clinical scales

Within the PD-PIGD group (n = 94), Spearman partial correlations, adjusted for age, revealed that PSV volume was associated with MMSE score (r = 0.25, p = 0.014) and MoCA score (r = 0.24, p = 0.02) (Table 3); however, these associations lost significance after further adjustment for sex and education (Supplementary Table 7). In contrast, significant age-adjusted Spearman partial correlations were found between higher PSV-L_PCL FC and lower PIGD score (r = −0.26, p = 0.013), and between higher PSV-L_PCL FC and higher MoCA score (r = 0.25, p = 0.015). Furthermore, these associations persisted even after additional adjustment for sex and education (PIGD: r = −0.26, p = 0.014; MoCA: r = 0.35, p < 0.001). Other significant age-adjusted correlations included negative associations between PIGD score and both PSV-R_PCG FC (r = −0.32, p = 0.002) and PSV-R_Put FC (r = −0.21, p = 0.04), and a positive association between PIGD score and PSV-R_Crus II FC (r = 0.25, p = 0.014); these associations also remained significant after additional adjustment for sex and education (Table 3 and Supplementary Table 7).

Table 3.

Spearman correlation analyses were used to assess the relationships between PSV metrics and clinical scales in patients with PD-PIGD

PSV metrics		PSV volume	PSV-L_PCL FC	PSV-R_PCG FC	PSV-L_FG FC	PSV-R_Put FC	PSV-R_Crus II FC
PIGD score	r	−0.17	−0.26	−0.32	0.19	−0.21	0.25
	p value	0.095	0.013	0.002	0.066	0.04	0.014
UPDRS III score	r	−0.10	0.02	−0.01	0.07	0.01	0.11
	p value	0.329	0.867	0.955	0.537	0.936	0.292
MMSE	r	0.25	0.21	0.19	−0.09	0.07	−0.13
	p value	0.014	0.044	0.065	0.377	0.526	0.216
MoCA	r	0.24	0.25	0.18	0.01	0.14	−0.22
	p value	0.02	0.015	0.089	0.893	0.173	0.037

Values represent Spearman partial correlation coefficients (r) and corresponding p-values, adjusted for age, within the PD-PIGD group (n = 94). Bold indicates statistical significance (p < 0.05).

L_FG left fusiform gyrus, L_PCL left paracentral lobule, MMSE Minimum Mental State Examination, MoCA Montreal Cognitive Assessment, PIGD postural instability and gait difficulty, PSV Posterior Superior Vermis, R_Crus II right cerebellar crus II, R_PCG right precentral gyrus, R_Put right putamen, UPDRS III Unified Parkinson’s Disease Rating Scale Part III.

Mediation analyses

To explore the interplay between PSV-L_PCL FC, cognitive function (MoCA score), and gait severity (PIGD score), we performed a mediation analysis, adjusting for potential confounders (age, sex, education, disease duration, LEDD, and UPDRS III score). Hierarchical linear regression established the prerequisite significant associations for testing a mediation model (Supplementary Tables 8–10). Subsequent bootstrap analysis (5000 samples) revealed a significant statistical indirect effect involving MoCA scores in the association between PSV-L_PCL FC and PIGD score (indirect effect: β = −2.58, 95% CI [−4.84, −0.32], p = 0.025). A significant direct association between PSV-L_PCL FC and PIGD score also remained (direct effect: β = −4.90, 95% CI [−9.80, −0.01], p = 0.049) (Table 4). These findings indicate that the relationship between higher PSV-L_PCL FC and better gait performance is partially mediated by better cognitive function.

Table 4.

Mediation effect of PSV-L_PCL FC on PIGD score via MoCA score

Effect Type	Path	β	p value	95% CI
Direct effect	PSV-L_PCL FC → PIGD score	−4.90	0.049	(−9.80, −0.01)
Indirect effect	PSV-L_PCL FC → MoCA → PIGD score	−2.58	0.025	(−4.84, −0.32)
Total effect	PSV-L_PCL FC → PIGD score	−7.48	0.004	(−12.52, −2.45)
Path coefficient	MoCA → PIGD score	−0.20	<0.001	(−0.31, −0.10)
	PSV-L_PCL FC → MoCA	12.62	0.005	(3.78, 21.46)

The mediation model was adjusted for age, sex, education years, disease duration, LEDD, and UPDRS-III score. Bold indicates statistical significance (p < 0.05). 95% CI based on bootstrapping (5000 samples).

CI confidence interval, LEDD levodopa equivalent daily dose, L_PCL left paracentral lobule, MoCA Montreal Cognitive Assessment, PIGD postural instability and gait difficulty, PSV posterior superior vermis, UPDRS III Unified Parkinson’s Disease Rating Scale Part III.

Discussion

This study provides stage-dependent evidence for regionally specific neurobiological alterations within the cerebellum of patients with the PD-PIGD phenotype. Our central finding is a non-linear, stage-dependent pattern of structural and functional alterations selectively involving the PSV. This pattern, characterized by a trend toward greater volume and robustly enhanced sensorimotor connectivity in early-stage disease (HY1), followed by distinct network reorganization in later stages (HY3–4), is consistent with a model of initial functional compensation that may transition to subsequent neurodegenerative failure. These findings distinguish the PSV from adjacent cerebellar areas and implicate it as a critical, stage-dependently vulnerable node in the pathophysiology of PD-PIGD.

The observed regional specificity highlights the unique role of the PSV as an early, targeted responder in PD-PIGD. Our analysis revealed that in HY1 patients, significant functional alterations were largely confined to PSV-cortical circuits, while the PIV and other cerebellar networks remained relatively quiescent. This finding supports a selective vulnerability or a highly specific adaptive response^8,21. We observed that, compared with HCs, HY1 patients demonstrated significantly enhanced PSV connectivity with key motor network nodes, particularly the sensorimotor cortical pathways including L_PCL and R_PCG, and this enhancement correlated with better clinical function. We interpret this as an adaptive compensatory mechanism, functionally plausible as the compromise of automatic motor control circuits within the basal ganglia may compel the brain to recruit alternative, top-down circuits for explicit, goal-directed motor control^22–24. This process is conceptually congruent with the scaffolding theory of aging and cognition (STAC), which posits that the brain recruits supplementary neural circuits to counteract functional decline²⁵. An analogous process is observed in healthy older adults, where demanding cognitive training elicits increased cerebellar–prefrontal connectivity that correlates with performance gains, supporting the cerebellum’s role in adaptive plasticity²⁶.

The biological substrate for this adaptive reorganization appears to be rooted in the dopamine-sensitive nature of the entire cerebello-thalamo-cortical pathway^6,27. The plausibility of recruiting this specific pathway is supported by its neurochemical architecture. Evidence from pharmacological fMRI shows that modulation of cerebello-cortical connectivity is spatially correlated with the density of underlying neurotransmitter receptors, including the dopaminergic D2 system²⁸. While the cerebellum itself has a relatively sparse dopaminergic innervation, the cortical targets of its output, such as the primary and supplementary motor areas, possess a significant density of D2 receptors critical for synaptic plasticity and motor learning^29,30. This D2-rich architecture at the cortical node renders the entire circuit highly sensitive to fluctuations in central dopaminergic tone, a functional relevance substantiated by findings that levodopa modulates this connectivity in patients^31–33. Underpinning this network-level adaptation, a confluence of cellular and systems-level processes may also contribute. At the synaptic level, the observed FC increase could be the macroscopic signature of plasticity, such as long-term potentiation (LTP)-like mechanisms³⁴. Concurrently, degeneration of key brainstem nuclei (e.g., the pedunculopontine nucleus or locus coeruleus) may trigger homeostatic plastic responses in their cerebellar targets, while non-neuronal responses like reactive astrogliosis could also contribute to the observed imaging signals^35–37. Although the relative contribution of these putative mechanisms cannot be disentangled with current imaging techniques, the positive correlation between PSV and cortical connectivity and preserved clinical function provides evidence that these multifaceted early changes collectively constitute an effective compensatory process.

In later disease stages (HY3–4), the neurobiological landscape appears to transition from effective compensation to maladaptive network dysfunction³⁸. Evidence for this shift first manifests as a failure of the primary compensatory pathway; we observed a significant weakening of FC between the PSV and the R_Put in HY3–4 patients relative to the HY1 group. This finding signifies a critical disruption of the cerebello-basal ganglia loop essential for motor control, a process likely compounded by progressive underlying pathology within the cerebellum itself, such as Purkinje cell loss^39,40. Concurrently, and in contrast to the focused and beneficial hyperconnectivity of the early stage, a topographically diffuse and functionally distinct pattern of reorganization emerges. We found a significant increase in PSV coupling with cognitive cerebellum (R_Crus II) and the fusiform gyrus, responsible for visual processing in these later stages⁴¹. This widespread, non-specific reorganization may represent a less efficient, and ultimately detrimental, secondary attempt at compensation, imposing an excessive attentional load on a cognitive control system already compromised by advancing PD^42,43. This state of “network overload”—where executive circuits are overburdened in managing tasks that should be automatic—is a leading candidate mechanism for the worsening of gait dysfunction and the emergence of phenomena like freezing of gait^44,45. Our own data substantiate this interpretation, revealing a significant positive correlation between greater PSV-R_Crus II FC and poorer gait performance. This entire process points to a model of exhausted plasticity and spreading neurodegeneration, as the maladaptive functional reorganization occurs alongside the atrophy trend of the cognitive cerebellum itself, suggesting these overburdened compensatory circuits are also undergoing structural failure^8,46. Therefore, the shift from a focused, functional compensation to a diffuse, interfering pattern does not simply represent a loss of function but an active, maladaptive process that contributes to clinical decline.

Our retrospective analysis in a PD-FoG cohort offers preliminary, yet functionally relevant, insights into the role of the PSV-L_PCL pathway in gait dysfunction, complementing the established literature implicating cerebellar circuits in PD pathophysiology^6,47. We found that targeting TMS closer to the individual’s peak FC locus within the L_PCL, known to be critical for lower limb sensorimotor control, correlated significantly with larger improvements following stimulation in specific gait tasks (SS180, 5mTUG). While this association did not extend to global motor scores (UPDRS III, PIGD) in this limited sample, the specificity to tasks, often disrupted in FoG, suggests that modulating this specific cerebello-cortical circuit may influence gait control mechanisms, particularly vulnerable in PD⁴⁸. This finding resonates with the growing rationale for utilizing FC to guide neuromodulation, aiming to enhance therapeutic precision by targeting patient-specific, functionally relevant neural circuits rather than relying solely on anatomical landmarks^49,50. Although large trials applying connectivity guidance in other conditions like depression have yielded complex results regarding superiority over standard protocols, our data provide disease-specific (PD-FoG) and symptom-specific (SS180, 5mTUG) correlative evidence supporting this principle^51,52. Furthermore, the observation that individual differences in baseline functional network architecture (PSV-L_PCL topography) relate to TMS outcome variability aligns conceptually with efforts to identify predictive biomarkers for therapeutic response in PD, analogous to how TMS-derived cortical excitability measures may predict response to dopaminergic therapy or how FC patterns may serve as disease biomarkers^53,54. Although the retrospective, correlational design and small PD-FoG cohort (n = 24) of this validation analysis preclude causal inference and limit generalizability, the findings lend valuable in vivo support to the biological relevance of our primary PSV-L_PCL FC results. Rigorous prospective, randomized controlled trials are now warranted to definitively assess the causality and clinical efficacy of connectivity-guided TMS targeting this pathway for gait dysfunction across broader PD populations, potentially incorporating baseline connectivity patterns for patient stratification.

Our finding that cognitive function (MoCA) partially mediates the statistical association between higher PSV-L_PCL FC and better gait performance (lower PIGD score) highlights the interplay between cerebellar circuits, cognition, and gait in PD-PIGD patients. This finding suggests that the PSV-L_PCL pathway influences gait both directly, likely reflecting sensorimotor contributions, and indirectly via its link with cognitive capacity^18,55. This aligns with established views that nonautomatic gait control in PD involves substantial cognitive resources, particularly executive functions and attention, likely engaging overlapping frontocerebellar networks involving the PSV⁴². The integrity of these interacting networks may be critical, as their compromise can underpin cognitive‒motor interference⁵⁶. However, caution is needed when interpreting these mediation findings. Given the study’s cross-sectional design, the model represents a statistical pattern of associations among PSV-L_PCL FC, cognitive function, and gait severity but does not provide evidence for a causal pathway. Establishing causality would require longitudinal data or experimental manipulation. Nevertheless, the observed partial mediation underscores that therapeutic strategies for PIGD might achieve greater efficacy by considering interventions that target both cerebellar circuit function and relevant cognitive domains.

Several limitations of this study warrant discussion. Primarily, the cross-sectional design fundamentally limits causal inference. The observed differences between HY stage groups represent group-level statistical comparisons at a single time point; therefore, we cannot definitively disentangle the temporal progression of neurobiological changes within individuals from potential confounders, such as cumulative medication effects. Although this design precludes the mapping of intra-individual trajectories, we sought to ground our primary FC findings in a clinically relevant context. Our retrospective validation in a separate PD-FoG cohort demonstrated that more accurately targeting the PSV-L_PCL pathway was associated with greater specific gait improvement after neuromodulation. This finding, while not establishing causality, provides preliminary imaging evidence for the functional relevance of this specific cerebello-cortical circuit to gait control in PD. Nevertheless, prospective longitudinal studies are essential to validate our proposed stage-dependent model of compensation-to-dysfunction, ideally focusing on de novo, drug-naive patients to preclude cumulative drug effects. Furthermore, our study has limitations regarding sampling method and motor subtype classification. Our cohort was enriched with patients exhibiting the PIGD phenotype, a focused strategy guided by our study goals and hypotheses. The current study relied on calculating the MDS-UPDRS ratio for TD/PIGD subtype differentiation and stratification. Although this approach is widely used for motor subtype classification, it inherently has some constraints. For example, it simplifies a continuous symptom spectrum into discrete categories and demonstrates temporal instability, with patients often transitioning between subtypes over time^3,57,58. This approach may also be incapable of fully capturing the considerable heterogeneity within the PIGD phenotype itself. Consequently, our findings should be interpreted as specific to the PIGD subtype, highlighting a set of stage-dependent alterations within this particular motor impairment pattern, rather than as a universal model applicable to other motor phenotypes like the tremor-dominant form. Finally, several other methodological factors require consideration. Our assessment of patients solely in the “off” medication state, while standardizing for disease stage, limits insights into the modulatory effects of dopaminergic therapy on cerebellar circuits. Methodologically, while SUIT-based normalization improves anatomical precision, the reliance on atlas-based regions cannot fully account for individual anatomical variability. The sensitivity of our volumetric findings to covariate adjustment also underscores the difficulty in isolating subtle structural changes from demographic factors in cross-sectional data. Our interpretations are thus subject to the inherent assumptions of rs-fMRI analyses and the resolution limits of the clinical scales employed.

In conclusion, this study characterizes the PSV as a site of regionally specific, nonlinear, and stage-dependent alterations in PD-PIGD phenotype, distinguishing its role from that of adjacent cerebellar areas. Early-stage (HY1) disease was marked by a potential compensatory signature, including a trend toward greater PSV volume (age-adjusted, though sensitive to further covariates) and enhanced FC between the PSV and L_PCL, which correlated with better gait and cognitive function. In later stages (HY3–4), this pattern shifted toward network dysfunction, evidenced by reduced PSV-R_Put connectivity and a trend toward atrophy in the cognitive cerebellum. The functional relevance of the early compensatory pathway was supported by a retrospective analysis linking targeted neuromodulation of this circuit to specific gait improvements. Furthermore, the interplay between these domains was highlighted by a partial mediation of the PSV-L_PCL FC-gait relationship by cognitive function. Taken together, these convergent findings delineate a unifying framework that implicates the PSV as a critical node in a trajectory from compensation to dysfunction in PD-PIGD. This positions the PSV as a promising candidate for investigation as a biomarker and a target for stage-specific therapeutic interventions, although its clinical utility should be confirmed through prospective and longitudinal studies.

Methods

Study design and participants

In this retrospective, cross-sectional study, the baseline data of patients were acquired between June 2018 and August 2024. The data included information from outpatients with PD treated at the First Affiliated Hospital of Anhui Medical University and data obtained during the baseline assessments of the participants enrolled in two independent clinical trials conducted at the same institution: ‘Repetitive Transcranial Magnetic Stimulation (rTMS) Treatment for Parkinson Disease’ (ClinicalTrials.gov: NCT02969941, registered on 2016-06-01) and ‘Effect of Theta-Burst Transcranial Magnetic Stimulation (TMS) for Freezing of Gait’ (ClinicalTrials.gov: NCT05192759, registered on 2021-11-22). In both trials, the recruited population consisted primarily of PD patients exhibiting gait impairments. A subset of baseline data from PD patients with freezing of gait (PD-FoG) who were enrolled in NCT05192759 was subsequently used for retrospective functional validation analysis, as detailed in the “Results” section^20,59. All procedures were approved by the Medical Ethics Committee of the First Affiliated Hospital of Anhui Medical University and conducted in accordance with the Helsinki Declaration, with written informed consent obtained from all participants.

Patients meeting the Movement Disorder Society (MDS) criteria for PD dementia were excluded⁶⁰. The initial cohort comprised 196 patients diagnosed with idiopathic PD according to the MDS clinical criteria⁶¹. Motor subtypes were determined using the MDS-Unified Parkinson’s Disease Rating Scale (UPDRS) ratio of TD/PIGD⁵⁸. Patients classified as PIGD (ratio ≤ 0.9, n = 108) were selected for the primary analysis, whereas TD (ratio ≥ 1.15, n = 46) and indeterminate (0.90 < ratio < 1.15, n = 42) subtypes were excluded. PD severity was staged using the Hoehn and Yahr (HY) scale⁶². Only PD-PIGD patients at HYs 1–4 were included, as HY-5 patients typically have significant mobility limitations preventing participation. All included PD patients were on stable dopaminergic medication for at least four weeks prior to assessment. Fifty-three age- and sex-matched HCs with no history of neurological or psychiatric disorders were recruited from the spouses of the patients with PD. After exclusions to ensure quality, the final primary analysis included 94 PD-PIGD patients (stratified as HY1, n = 33; HY2, n = 30; HY3–4, n = 31) and 46 HCs (Fig. 1).

Clinical assessments

Baseline clinical assessments and MRI scans for this analysis were performed prior to any trial-related TMS intervention and while participants were in a practically defined ‘off’ medication state (≥12-h withdrawal of dopaminergic drugs), thereby minimizing acute medication effects and standardizing disease stage comparisons^33,63. Motor function was evaluated using the MDS-UPDRS⁶⁴. Key measures included the total score of the motor examination (Part III, UPDRS-III) and the PIGD score, which was calculated as the sum of items from both Part II (2.12, 2.13) and Part III (3.10, 3.11, 3.12)⁵⁸. Global cognitive function was assessed with the Montreal Cognitive Assessment (MoCA) and Mini-Mental State Examination (MMSE). Emotional symptoms were measured using the Hamilton Anxiety Scale (HAMA) and the Hamilton Depression Scale (HAMD). Demographic information (age, sex, and education years) and disease duration were recorded. The levodopa equivalent daily dose (LEDD) was calculated for all PD-PIGD patients⁶⁵.

MRI data acquisition

MRI data were acquired on a 3.0 T GE Discovery 750 scanner (GE Healthcare, Milwaukee, WI, USA) using an 8-channel head coil at the University of Science and Technology of China. High-resolution T1-weighted anatomical images were acquired using a 3D BRAVO sequence with the following parameters: repetition time (TR)/echo time (TE) = 8.16/3.18 ms; inversion time = 450 ms; flip angle = 12°; field of view (FOV) = 256 × 256 mm²; matrix = 256 × 256; slice thickness = 1 mm (no gap); 188 sagittal slices; voxel size = 1 × 1 × 1 mm³. Resting-state functional images were acquired with a gradient-echo echo-planar imaging (EPI) sequence: TR/TE = 2400/30 ms; flip angle = 90°; FOV = 192 × 192 mm²; matrix = 64 × 64; slice thickness = 3 mm (no gap); and 46 transverse slices parallel to the anterior‒posterior commissure (AC-PC) line, 217 volumes. During scanning, the participants were instructed to remain awake with their eyes closed and minimize head movement.

MRI quality control and participant exclusion

Participants were excluded if visual inspection of structural MR images revealed neurological comorbidities (e.g., focal lesions, tumors, severe white matter disease), if rs-fMRI data revealed excessive head motion (maximum absolute displacement >3 mm translation or >3° rotation relative to the initial volume), or if other significant artifacts were present⁶⁶. These criteria led to the exclusion of 14 PD-PIGD patients and 7 HCs. For the remaining participants included in the final analyses, there were no missing data for the variables examined.

MRI preprocessing and denoising

Standard rs-fMRI data processing was performed using the CONN toolbox (v.22a) implemented in Statistical Parametric Mapping (SPM12)⁶⁷. The preprocessing pipeline included the following steps: discarding the first 10 volumes for signal equilibrium; slice-timing correction; realignment and unwarping to correct for motion and susceptibility-distortion interactions; outlier identification using Artifact Detection Tools (ART), where scans with framewise displacement >0.9 mm or global BOLD signal changes >5 standard deviations were flagged; direct segmentation and spatial normalization into Montreal Neurological Institute (MNI) space (resampled to 3 × 3 × 3 mm³ voxels) using the unified segmentation algorithm; and spatial smoothing with a 6-mm full-width at half-maximum (FWHM) Gaussian kernel^66,68. Subsequently, a comprehensive denoising pipeline was implemented. This involved a single linear regression model that included several nuisance covariates: the 12 motion parameters derived from realignment, the flagged outlier time points (scrubbing), and five principal components each from white matter and cerebrospinal fluid signals, estimated via the anatomical CompCor (aCompCor) method⁶⁹. After regression of these covariates, the residual BOLD time series was temporally band-pass filtered between 0.009 and 0.08 Hz.

Structural MRI analysis

Cerebellar gray matter (GM) volume was analyzed with the spatially unbiased infratentorial toolbox (SUIT v3.3) in SPM12⁷⁰. T1-weighted images underwent automated segmentation, generating GM, white matter, and cerebrospinal fluid probability maps, with simultaneous isolation of the cerebellum with SUIT routines. For more accurate alignment, individual cerebellums were nonlinearly registered to the high-resolution SUIT atlas template (1 mm³ isotropic) employing the Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) algorithm, yielding participant-specific deformation fields^71,72. These fields were subsequently applied to warp the native-space GM probability maps into the standard SUIT template space. Critically, this spatial normalization incorporated modulation by Jacobian determinants derived from the deformation fields to preserve local GM volume information during warping. Following normalization, the quality of individual cerebellar segmentation and alignment was visually inspected by two trained raters who were blinded to the participant groups.

Volumetric quantification for individual SUIT atlas parcels (e.g., Vermis VI) involved summing modulated GM probability values within the parcel mask multiplied by the voxel volume (1 mm³). Volumes were quantified for composite regions of interest (ROIs) defined as follows: PSV (Vermis VI, Crus I, Crus II, and VIIb) and PIV (Vermis VIIIa, VIIIb, IX, and X), as depicted in Fig. 2a. In addition, based on established anatomical and functional parcellations of the cerebellum, we analyzed other subregions including the anterior lobe (lobules I–IV), cognitive cerebellum (CBMc: bilateral Crus I/II), motor cerebellum (CBMm: bilateral lobules V, VI, VIIb, and VIIIa/b), and posterior inferior cerebellum (lobules IX and X)^39,55. These subregions were selected on the basis of their anatomical location and known functions (motor/cognitive), aiming to provide a comprehensive comparative framework (detailed ROI definitions in Supplementary Table 1). Given that the ‘modulation’ step in SUIT, which scales voxelwise GM probability by the Jacobian determinants of the deformation fields, inherently corrects for global brain size differences relative to the template space, total intracranial volume was not included as an explicit covariate in subsequent statistical models, consistent with prior studies employing similar approaches^8,39,72.

Functional MRI analysis

Seed-based whole-brain FC analyses were performed via the CONN toolbox. Seed ROIs, defined in MNI space on the basis of the SUIT atlas, included the entire PV (PSV + PIV), PSV, PIV, Lobules I-IV, CBMc, CBMm, Lobules IX, and X. For each participant and each seed ROI, the average BOLD time series was extracted from all voxels within the seed mask. First-level analyses computed Pearson correlation coefficients between each seed’s time series and the time series of all other brain voxels, followed by Fisher’s r-to-z transformation to generate individual whole-brain FC maps. At the second level, we first conducted an exploratory analysis comparing the overall PD-PIGD group with the HC group using the PV seed. The primary analysis then involved voxelwise general linear models (GLMs) to compare FC maps across the four groups (HC, HY1, HY2, and HY3–4) for each of the specified cerebellar seed ROIs.

Statistical significance for group comparisons was determined using a primary voxel-level threshold of p < 0.001 (uncorrected) combined with a cluster-level threshold of p < 0.05 corrected for multiple comparisons using the false discovery rate (FDR) method⁷³. Significant clusters identified in group analyses were first anatomically localized using the automated anatomical labeling (AAL) atlas⁷⁴. Subsequently, cortical clusters were functionally assigned to large-scale resting-state networks based on the Yeo 7-network parcellation⁷⁵. Cerebellar clusters were attributed to a ‘Cerebellum Network’ classification, and subcortical clusters (e.g., putamen) were classified as ‘Subcortical’, reflecting their anatomical identification and broader functional systems. For subsequent correlation and mediation analyses, mean z-transformed FC values were extracted from the significant clusters identified in the four-group PSV-seed GLM analysis.

Statistical analysis

Statistical analyses for demographic and clinical data were performed using JASP (v0.19.1). All statistical tests were two-tailed, with a significance level set at α = 0.05. Demographic and clinical characteristics were compared across the four groups (HC, HY1, HY2, and HY3-4) using one-way analysis of variance (ANOVA) for continuous variables and chi-square tests for categorical variables. Assumptions for parametric tests (i.e., normality of distributions and homogeneity of variances) were assessed to ensure the validity of the selected statistical methods. Post hoc pairwise comparisons following a significant ANOVA were adjusted for multiple comparisons using the Bonferroni method. Continuous data are presented as mean ± standard deviation (SD) and range (minimum‒maximum), while categorical data are presented as counts (n).

In the primary analyses of group differences, regional cerebellar volumes were assessed using analysis of covariance (ANCOVA), whereas group differences in FC were assessed via voxelwise GLMs. In both sets of analyses, age was included as a primary covariate, given the significant group difference in age and its known impact on brain structure and function^72,76. To confirm the robustness of the primary findings, sensitivity analyses were conducted for all main group comparisons by additionally adjusting for sex and education years. Within the PD-PIGD group, associations between primary imaging metrics (PSV volume; mean FC values from significant PSV-seed clusters) and clinical scores (PIGD, UPDRS-III, MMSE, MoCA) were examined using partial Spearman’s rank correlation, with sensitivity analyses including additional adjustments for age, sex, and education.

To investigate cognitive function as a potential mediator of the relationship between significant PSV metrics and PIGD score, we performed mediation analysis. Prerequisite pathway associations (predictor-outcome, predictor-mediator, and mediator-outcome adjusted for predictor) were first established using hierarchical linear regression and adjusted for age, sex, education years, disease duration, LEDD, and UPDRS III score. The absence of problematic multicollinearity among the predictors was confirmed (variance inflation factors < 3.0). If prerequisites were met, mediation effects were estimated by bootstrapping (5000 samples) to derive 95% confidence intervals (CI) for the direct and indirect pathways.

Author contributions

L.-Z.-X.Y., R.Y., and P.-P.H. were primarily responsible for the study design. L.-Z.-X.Y. wrote the first draft; J.-Y.H. and X.C. analyzed the clinical data and statistical analysis; L.-L.H., M.-Q.W., M.-H.Z., J.-J.C., P.-P.L. and L.F. carried out MRI data collection; Y.-Q.L., J.-J.W. and X.-Y.Z. contributed to patient management and sample collection; J.-M.S. and G.-J.J. analyzed the MRI data; K.W., R.Y., and P.-P.H. revised the paper. All authors read and approved the final version of the paper.

Data availability

The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41531-025-01065-1.