From June 2024 to May 2025, 46 PD patients were recruited from Shenzhen People’s Hospital to participate in this cross-sectional study. Inclusion required: (1) probable PD diagnosis based on the 2015 Movement Disorder Society (MDS) criteria [14], and (2) reduced striatal ¹⁸F-FP-DTBZ uptake. Exclusions were major systemic diseases, cognitive assessment incompatibility, cerebrovascular/atrophic changes, magnetic resonance (MR) contraindications, and claustrophobia. Clinical assessments, ¹⁸F-FP-DTBZ PET, and rs-fMRI were performed on all individuals. Neurologists who were blind to the participants’ condition collected demographics, levodopa equivalent daily dose (LEDD) data, disease duration, and OFF-state Unified Parkinson’s Disease Rating Scale Part III (UPDRS-III) scores [15].

Cognitive examinations were done in the ON-state to minimize motor symptom interference. SCD was defined as an SCD-Scale-9 score $\ge$3 [16]. Objective cognitive function was measured by standardized tests [6], including the Verbal Fluency Test-Alternating (VFT), Boston Naming Test (30-item version, BNT), Loewenstein-Acevedo Scales for Semantic Interference and Learning (LASSI-L), Digit Span Test (DST), Symbol Digit Modalities Test (SDMT), Judgement of Line Orientation (JLO), Shape Trails Test (STT), and Stroop Color and Word Test (SCWT). Depressive and anxiety symptoms were assessed using the Hamilton Depression Scale (HAMD) and Hamilton Anxiety Scale (HAMA), respectively.

Diagnostic classifications followed established criteria: PD-MCI: impairment on $\ge$2 neuropsychological tests according to MDS Level II criteria (2012) [6], with preserved activities of daily living; PD-SCD: SCD-Scale-9 score $\ge$3 without meeting PD-MCI criteria [16]; PD-NC: absence of SCD (SCD-Scale-9 score 3) and normal objective cognition.

All imaging was performed on a 3.0 T PET/MR scanner (uPMR 790, United Imaging, Shanghai, China) with a 24-channel head coil. To minimize acute medication effects, all participants were required to discontinue their anti-Parkinson’s medications for at least 12 h before undergoing the ¹⁸F-FP-DTBZ PET scan. All participants underwent simultaneous PET and MRI scanning. PET scanning commenced 90 min post-intravenous ¹⁸F-FP-DTBZ injection, with participants positioned supine and instructed to remain still throughout the scan [17]. Images were reconstructed using ordered subset expectation maximization (OSEM) with time-of-flight (TOF), and Dixon-based MR attenuation correction (acquisition: 10 min; matrix: 192 $\times$ 192; slice thickness: 2 mm; iterations/subsets: 4/20; full width at half maximum (FWHM): 3.0 mm). Simultaneous MR imaging included both structural and functional sequences. Resting-state BOLD-fMRI was acquired using a gradient-echo echo-planar imaging (EPI) sequence with the following parameters: repetition time (TR) = 2000 ms; echo time (TE) = 30 m; field of view (FOV) = 224 $\times$ 224 mm²; matrix = 64 $\times$ 64; slice thickness = 3.5 mm; and slice gap = 0.7 mm. High-resolution T1-weighted structural images were obtained with the following parameters: TR = 7.8 ms; TE = 3 ms; flip angle (FA) = 10^∘; FOV = 240 $\times$ 240 mm²; matrix = 240 $\times$ 240; and slice thickness = 1 mm. Participants were told to close their eyes, stay awake, and avoid head motion during scanning. Standard hybrid PET/MR protocols ensured accurate co-registration and quantitation.

PET images were preprocessed with the Statistical Parametric Mapping 12 (SPM12) toolbox implemented in MATLAB R2023b (Wellcome Trust Centre for Neuroimaging, London, UK; https://www.fil.ion.ucl.ac.uk/spm) [17]. Initially, for each subject, the PET image was rigidly coregistered to the corresponding T1-weighted anatomical MRI using normalized mutual information. Subsequently, tissue segmentation of the T1-weighted image was performed to generate probabilistic maps of gray matter, white matter, and cerebrospinal fluid, which were also utilized for partial volume effect (PVE) correction. The coregistered PET images were then normalized to the Montreal Neurological Institute (MNI) standard space with isotropic voxels of 2 mm using the forward deformation fields derived from tissue segmentation. Finally, the normalized PET images were smoothed using an isotropic Gaussian kernel with an 8 mm FWHM. Preprocessed PET images were coregistered to the Automated Anatomical Labeling (AAL 116) template for region-based analysis. ¹⁸F-FP-DTBZ binding ratios were calculated using the following formula, based on standardized uptake value (SUV):

$$\left[{}^{18}\mathrm{F}\right]\mathrm{FP}-\mathrm{DTBZ}\text{binding ratio}=\frac{{\mathrm{SUV}}_{\text{striatal mean}}-{\mathrm{SUV}}_{\text{occipital mean}}}{{\mathrm{SUV}}_{\text{occipital mean}}}$$

Striatal subregions included the caudate nucleus and putamen. The occipital reference region comprised all occipital lobe areas in the AAL 116 atlas, including the calcarine, cuneus, lingual gyrus, superior occipital gyrus, middle occipital gyrus, and inferior occipital gyrus. Regions exhibiting significant group differences in ¹⁸F-FP-DTBZ binding were defined as regions of interest (ROIs).

Rs-fMRI images were preprocessed with the Data Processing & Analysis for Brain Imaging toolbox (DPABI V8.2, http://www.rfmri.org/dpabi). The first 10 time points were discarded to allow for signal equilibration and participant habituation. Slice timing correction was applied, followed by realignment for head motion correction. Nuisance covariates, including 24 head motion parameters, white matter, cerebrospinal fluid, and global mean signals, were regressed out of the time series. The normalized functional images were then spatially smoothed with a 4-mm FWHM Gaussian kernel. Finally, a temporal band-pass filter (0.01–0.1 Hz) was applied to the time series to retain low-frequency fluctuations relevant for resting-state connectivity analysis. Participant data with excessive head motion (more than 30% of frames with framewise displacement exceeding 0.3 mm) were excluded from further analyses.

SFC quantifies direct and indirect (multi-step) brain area functional interactions using fMRI-derived functional connectivity matrices [12, 13]. After fMRI data preprocessing, a whole-brain functional connectivity matrix was constructed for each participant by calculating the Pearson correlation coefficients between the BOLD time series of all brain region pairs. Only positive connections that were statistically significant (false discovery rate (FDR)-corrected with q 0.001) were retained, and binarized matrices were used to identify important connections. The striatal regions exhibiting significant group differences in ¹⁸F-FP-DTBZ binding ratios were selected as seed regions. For each participant, SFC matrices were generated by counting, for each brain region, the number of unique paths of exact step length l that connect the seed region to the target node (where l = 1 represents direct connectivity, and l = 2–7 represents indirect multi-step connections; after 7 steps, SFC patterns tend to reach a stable state). Specifically, at any step-distance l, the SFC degree of a target node i is defined as the total number of unique paths of length l connecting i to the seed. In the current study, the findwalks.m function from the Brain Connectivity Toolbox was used to compute SFC for each participant [18, 19]. At each step, the SFC matrix was standardized using a z-transformation: for each region, the SFC value was centered by subtracting the whole-brain mean SFC and dividing by its standard deviation at that step. This normalization provided a measure of the relative increase or decrease of connectivity degree under a given step-distance. Multi-step connections were grouped as short-range (2–3), medium-range (4–5), and long-range (6–7); only odd-numbered steps were reported to reduce redundancy.

In line with previous studies [20], the stable step was calculated for each participant to quantify the convergence rate of SFC patterns. The stable step was defined as the minimum step k at which the Pearson correlation coefficient between SFC spatial maps remained $\ge$0.999 for three consecutive steps [r(k, k + 1), r(k + 1, k + 2), and r(k + 2, k + 3) all $\ge$0.999], indicating pattern stabilization.

To interpret SFC patterns within the context of canonical cortical functional networks, we quantified the overlap between the SFC maps and the seven intrinsic networks defined by the Yeo-7 atlas [21]. Specifically, for each group-level SFC map at multiple-step distances, the SFC image derived from the AAL template was resampled to match the spatial resolution and orientation of the Yeo-7 network template. During this process, each AAL-based SFC image was spatially aligned to the Yeo-7 template, and nearest-neighbor interpolation was used to preserve the discrete labeling of network regions. For each step and each group, mean SFC values were computed separately within each of the seven canonical Yeo-7 network masks. To facilitate group- and step-wise comparison of hierarchical network involvement, the mean SFC values for each network and each step were visualized as heatmaps. These visualizations enabled a systematic comparison of SFC pattern evolution across cortical networks and between diagnostic groups.

Factor analysis is used to reduce the number of clinical indicators and extract key features. The suitability of the dataset for factor analysis was first evaluated using the Kaiser-Meyer-Olkin (KMO) measure and Bartlett’s test of sphericity. Only datasets meeting standard thresholds (KMO $>$0.6 and Bartlett’s test p 0.05) proceeded to further multivariate analysis [22]. To determine the optimal number of latent factors, a scree plot and parallel analysis were performed (as Supplementary Fig. 1), and the eigenvalues and explained variance of each factor were calculated. Exploratory factor analysis was then conducted using the principal axis factoring method with oblimin (oblique) rotation to allow for correlations among the factors. The magnitude and pattern of the factor loadings were used to characterize the underlying clinical cognitive dimensions.

Differences among groups in demographics and neuropsychology were assessed using one-way analysis of variance (ANOVA), Kruskal-Wallis tests, or $\chi$² tests as appropriate. Comparing ¹⁸F-FP-DTBZ binding ratios and SFC values between groups were performed with analysis of covariance (ANCOVA), using sex, age, and education as covariates, and FDR-correction for multiple comparisons (significance threshold: p 0.05). Effect sizes were calculated using Cohen’s d. Covariate-adjusted partial correlation analyses were used to examine the relationships between SFC metrics and caudate ¹⁸F-FP-DTBZ binding ratios, as well as cognitive factors, respectively. All analyses were performed in MATLAB R2023b (MathWorks, Inc., Natick, MA, USA) and R 4.2.0 (R Foundation for Statistical Computing, Vienna, Austria), with visualizations generated using the BrainNet Viewer toolbox (V1.7, https://www.nitrc.org/projects/bnv/).

This study included 46 PD patients divided into three cognitive subgroups: PD-NC (n = 15), PD-SCD (n = 16), and PD-MCI (n = 15). As shown in Table 1, the groups exhibited comparable demographics (age, sex, education) and clinical parameters (disease duration, levodopa dosage; all p $>$ 0.05). However, the PD-MCI group demonstrated significantly more severe motor deficits (UPDRS-III: PD-MCI, 35.6 vs. PD-NC, 24.9, and PD-SCD, 24.4; p = 0.013).

Neuropsychological assessments revealed significant cognitive differences in all five domains (PD-MCI vs. PD-SCD/PD-NC; all p $\le$ 0.001). Executive dysfunction in PD-MCI was characterized by prolonged STT completion times (400.0 s vs. 237.0/174.0 s) and increased SCWT interference (34.0 s vs. 30.0/20.0 s). PD-MCI patients exhibited pathological LASSI-L memory patterns, including exaggerated proactive semantic interference (B1 recall: 2.0 vs. 5.0/6.0), impaired interference recovery (B2 recall: 4.0 vs. 8.0/10.0), and amplified retroactive semantic interference (A3 recall: 3.0 vs. 5.0/5.0). Language deficits were evident in reduced VFT (7.0 vs. 11.0/15.0) and BNT scores (16.0 vs. 22.5/25.0). Attention and working memory impairments (DST: 9.0 vs. 11.0/12.0; SDMT: 16.0 vs. 26.5/31.0) and visuospatial deficits (JLO: 15.0 vs. 20.0/24.0) further delineated the multidomain cognitive profile of PD-MCI. No statistically significant differences were observed in HAMD and HAMA scores among the three groups.

Quantitative analysis of the ¹⁸F-FP-DTBZ specific uptake ratio in the dorsal striatum revealed significant differences in caudate nucleus VMAT2 availability across cognitive subgroups of PD (Table 2). PD-MCI patients had significantly lower caudate nucleus VMAT2 activity than PD-NC and PD-SCD cohorts (left caudate: ANCOVA FDR p = 0.0320; right caudate: ANCOVA FDR p = 0.0357, Supplementary Fig. 2). We also observed increased VMAT2 activity in the bilateral caudate nucleus in patients with PD-SCD. In contrast, putamen showed no significant intergroup variations. These findings supported the selection of the caudate nucleus as the seed region for SFC analyses. The left and right caudate masks were merged to form a single bilateral composite seed for all SFC analyses.

SFC analysis revealed progressive alterations in caudate-centered network across the PD cognitive subgroups (Fig. 1). In PD-NC, the caudate nucleus maintained stable moderate-strength connections with the prefrontal (orbitofrontal cortex, OFC; inferior frontal gyrus, IFG), motor (supplementary motor area, SMA), subcortical (lentiform nucleus), and limbic regions. As connection steps increased, connectivity with temporal lobe decreased, whereas the thalamic connectivity strengthened at longer path distances, indicating preserved network adaptability. In PD-SCD, direct connectivity between the caudate and OFC was lower than in PD-NC. Nevertheless, widespread hyperconnectivity with limbic and frontal regions emerged, reflecting early compensatory reorganization of brain networks. In contrast, PD-MCI exhibited globally attenuated connectivity, with notable disconnection between the caudate and SMA, as well as disrupted distance-dependent network organization. Collectively, this delineated a triphasic trajectory of network dysfunction in early PD-CI: stable connectivity in PD-NC developed into hyperconnected reorganization in PD-SCD, then progressed to broad-scale network inefficiency in PD-MCI.

Fig. 1.

SFC strength for odd-step distances in PD-NC, PD-SCD, and PD-MCI groups. Color bar: Represents SFC value, ranging from 0 to 1. The color scale transitions from blue to red, encode the increasing magnitude of SFC value. SFC, stepwise functional connectivity.

To further contextualize caudate-centered network dynamics in early cognitive progression of PD, we mapped the stepwise functional connectivity of the caudate nucleus onto the Yeo-7 networks (Fig. 2). PD-NC showed moderate caudate overlap with several functional networks, including the frontoparietal, somatomotor, salience, limbic, and default mode networks—with spatial overlap ranging from 11% to 39%. PD-SCD exhibited marked hyperintegration (up to 59%), particularly in the salience, frontoparietal, and somatomotor networks. In contrast, PD-MCI showed reduced and scattered integration (2%–30%) across all networks.

Fig. 2.

Heatmap quantification of spatial overlap between SFC odd-step maps and Yeo-7 cortical networks across PD subtypes. Matrix rows represent SFC step lengths (1, 3, 5, 7), and columns denote the seven canonical networks. Cell color intensity and numeric values indicate the percentage of overlapping voxels relative to the total network volume, and warm colors indicate high spatial overlap.

Fig. 3 illustrates significant intergroup differences in caudate-centered stepwise functional connectivity (FDR, p 0.05). In direct pathways, PD-SCD patients exhibited lower caudate–right superior orbitofrontal (ORBsup) connectivity but greater caudate–thalamic connectivity than did PD-NC. Moreover, SFC analysis from step 3 to step 7 revealed that progressive hyperconnectivity emerged between the caudate and the SMA, right superior frontal gyrus (SFG), right hippocampus, and globus pallidus (GP). PD-SCD also exhibited transient hyperconnectivity in the right middle cingulate cortex (MCC, step 3) and right middle frontal gyrus (MFG, step 5). Mid-to-long-range connections (steps 5–7) in PD-SCD further showed increased coupling with the left inferior temporal gyrus (ITG).

Fig. 3.

Intergroup differences in SFC strength for odd-step distances. Top to bottom: Pairwise comparisons of PD cognitive subgroups (PD-SCD vs. PD-NC; PD-MCI vs. PD-NC; PD-MCI vs. PD-SCD). Group differences were assessed using an analysis of covariance adjusted for age, sex, and years of education. All results survive FDR correction at p 0.05. Color bar: Represents Cohen’s d effect sizes, ranging from –1 to 1. Warm colors (red-yellow): Higher SFC connectivity strength in the first-named group (e.g., PD-SCD in the top panel). Cool colors (blue-cyan): Higher SFC connectivity strength in the second-named group (e.g., PD-NC in the top panel).

PD-MCI had fewer direct caudate-thalamus and caudate-right ORBsup connections than did PD-NC. From step 3 to step 7, progressive disengagement was observed in the connections between the caudate and the SMA, right hippocampus, and GP. Mid-to-long-range connections (steps 5–7) demonstrated decreased left ITG integration. Right MCC (step 3) and MFG (step 5) showed phase-specific decreases, whereas the right SFG exhibited initial hyperconnectivity at step 3, followed by progressive attenuation at step 7.

The intergroup analysis between PD-MCI and PD-SCD revealed that the patterns of functional connectivity alterations were largely consistent with those identified in the PD-MCI versus PD-NC comparisons. Two notable exceptions were identified: PD-MCI exhibited more accelerated disconnection of the right SFG at step 3 and attenuation of right IFG connectivity at step 5.

Stable steps for caudate nucleus connections varied significantly among PD early cognitive subgroups (Fig. 4). PD-MCI patients had longer stable steps than did PD-NC and PD-SCD (ANCOVA; p 0.05). This delayed stabilization of PD-MCI reflected rigidified functional connection patterns and limited adaptive state transitions, indicating impaired dynamic network reconfiguration capacity. Notably, PD-NC and PD-SCD showed similar stable steps, indicating preserved neural flexibility in preclinical stages.

Fig. 4.

Group comparisons of stable step. Differences in stable steps among PD cognitive subgroups were assessed using analysis of covariance adjusted for age, sex, and years of education. * indicates statistically significant differences (FDR-corrected p 0.05).

Partial correlation analyses were performed to assess metabolic-functional coupling between caudate nucleus ¹⁸F-FP-DTBZ uptake ratios and SFC metrics (Table 3). We observed a significant negative correlation of stable steps with left caudate uptake (r = –0.366, p = 0.019). Furthermore, bilateral caudate ¹⁸F-FP-DTBZ uptake demonstrated positive correlations with multi-step SFC values in the right hippocampus and GP (all p 0.05). These findings collectively suggested robust coupling between dopaminergic metabolism and functional connectivity.

Factor analysis addressed multicollinearity among the 10 neuropsychological parameters, extracting two oblique cognitive factors that collectively explained 87.75% of the variation (Table 4). Factor 1 (global cognitive efficiency; 78.15% variance) integrated high-loading scores from five cognitive domains: language (VFT, BNT), attention/working memory (DST, SDMT), visuospatial processing (JLO), memory (LASSI-L A3), and executive control (STT, SCWT). Factor 2 (memory control; 9.6% variance) was primarily driven by LASSI-L indices of proactive semantic interference and recovery, reflecting inhibitory efficiency during memory encoding and retrieval.

Partial correlation analyses were performed to assess the relationships between SFC values of each brain region at each step and cognitive factor scores (Fig. 5). Global cognitive efficiency (Factor 1) was positively correlated with left GP connectivity at step 3 (r = 0.3156, p = 0.0422) and right ORBsup connectivity at step 7 (r = 0.3244, p = 0.0366), suggesting that striatal-OFC networks integration mediate advanced cognitive operations across early processing through late-stage coordination. Memory control (Factor 2) demonstrated robust associations with the left ITG at step 5 (r = 0.3738, p = 0.0153) and step 7 (r = 0.3503, p = 0.0235), indicating that caudate-ITG circuit engagement supports interference-resistant memory during mid-to-long-range connections.

Fig. 5.

Partial correlations between cognitive factors and SFC values adjusted for age, sex, and years of education. (A,B) Factor 1 correlations with SFC values at (A) Step 3 left-GP and (B) Step 7 right-ORBsup. (C,D) Factor 2 correlations with SFC values at (C) Step 5 and (D) Step 7 left ITG. GP, globus pallidus; ORBsup, superior orbitofrontal cortex; ITG, inferior temporal gyrus.