Participants
This study integrated data from three neuroimaging databases: the Parkinson’s Progression Markers Initiative (PPMI), the Health and Aging Brain Study–Health Disparities (HABS), and the Alzheimer’s Disease Neuroimaging Initiative (ADNI). All participants provided written informed consent, and all procedures were conducted in accordance with the Declaration of Helsinki.
From the PPMI database, 95 individuals with a diagnosis of Parkinson’s disease (PD) were included. All PD participants met the PPMI eligibility criteria: diagnosis of PD within the prior two years, absence of dementia, Hoehn & Yahr stage ≤ 2, and drug-naïve or minimally treated status at enrollment. Participants were enrolled at baseline and followed longitudinally across three imaging sessions (baseline, approximately 2 years, and approximately 4 years), yielding a total of 285 PD observations.
Healthy control (HC) participants (n = 50) were drawn from two sources: (1) the HABS dataset (n = 30), a community-based study of brain aging and health disparities that includes cognitively unimpaired older adults recruited across diverse ethnic and racial groups32; and (2) the ADNI dataset (n = 20), which provides longitudinal neuroimaging data from cognitively normal older adults alongside individuals with mild cognitive impairment and Alzheimer’s disease33. HC participants were selected on the basis of: absence of neurological or psychiatric diagnoses, no significant cognitive impairment (MoCA ≥ 26 or equivalent), and availability of compatible 3 T diffusion-weighted MRI sequences suitable for DTI-ALPS computation. HC participants contributed baseline and 3-year follow-up imaging data.
Demographics and clinical assessments
Demographic variables included age, sex, ethnicity, and education. Clinical data comprised disease duration, baseline motor features (bradykinesia, tremor, rigidity, postural instability), and cognitive performance. Cognitive performance was assessed with the Montreal Cognitive Assessment (MoCA)34, supplemented by domain-specific tests including Hopkins Verbal Learning Tests (Total (HVLTR) and delayed (HVLDR)) for memory35, Symbol Digit Modalities Test (SDM) for processing speed36, Benton Judgment of Line Orientation (JLO) for visuospatial function37, Letter–Number Sequencing (LNS) for working memory38, and Semantic Fluency (SF) for semantic verbal fluency39. Depressive symptoms were evaluated with the short Geriatric Depression Scale40. Probable REM sleep behavior disorder (pRBD) was screened with the 13-item RBD Screening Questionnaire (RBDSQ) with a cutoff score ≥ 5 as positive41.
MRi acquisitionPD cohort
Structural MRI data were acquired on a 3.0-T Siemens TrioTim scanner using a 3D T1-weighted MPRAGE sequence with the following parameters: TR = 2300 ms, TE = 2.98 ms, inversion time (TI) = 900 ms, flip angle = 9°, 176 slices, matrix size = 256 × 256, and voxel size = 0.976 × 0.976 × 1 mm³. Parallel imaging was performed using GRAPPA with an acceleration factor of 2.
Diffusion MRI data were acquired using a single-shot 2D echo-planar imaging sequence with monopolar diffusion encoding: TR = 3800 ms, TE = 88 ms, flip angle = 90°, 72 slices, matrix size = 116 × 116, voxel size = 2 × 2 × 2 mm³, GRAPPA acceleration factor of 2, no multiband acceleration. The diffusion protocol included 64 diffusion-weighted directions (b = 1000 s/mm²) and one non-diffusion-weighted (b = 0 s/mm²) image.
Healthy controls
Structural MRI data for HABS participants were acquired on a 3 T Siemens scanner using a 3D T1-weighted MPRAGE sequence with whole-brain coverage (matrix size: 240 × 256 × 176; voxel size: 1.05 × 1.05 × 1.20 mm³, TR = 2300 ms; TE = 2.95 ms; TI = 900 ms; flip angle = 9°). Diffusion MRI data were acquired on a 3 T Siemens scanner using a single-shot spin-echo echo-planar imaging sequence (TR ≈ 8000 ms; TE ≈ 89 ms; flip angle = 90°) with isotropic voxel dimensions of 2 × 2 × 2 mm³ (matrix size ≈ 128 × 128; ~70 axial slices; FOV ≈ 256 × 256 mm). The acquisition included 35 volumes consisting of diffusion-weighted images acquired along 30 gradient directions at b = 1000 s/mm² and 5 non-diffusion-weighted (b = 0) volumes32. Structural MRI data for ADNI participants were acquired on 3 T Siemens Magnetom Prisma scanners following the ADNI3 neuroimaging protocol, using a 3D T1-weighted MPRAGE sequence with the following parameters: TR = 2300 ms, TE = 2.98 ms, TI = 900 ms, flip angle = 9°, 176 slices, and voxel size = 1.0 × 1.0 × 1.0 mm isotropic, with GRAPPA acceleration factor of 2. Diffusion MRI data were acquired using a single-shot 2D EPI sequence with the following parameters: TR = 3300 ms, TE = 71 ms, voxel size = 2.0 × 2.0 × 2.0 mm isotropic, comprising 48 diffusion-weighted directions at b = 1000 s/mm² and 7 non-diffusion-weighted (b = 0) volumes33.
MRI processing
Structural and diffusion MRI data were processed using micapipe (v0.2.3), integrating AFNI, FSL (v6.0.5), ANTs (v2.3.5), MRtrix3 (v3.0.3), and FreeSurfer (v7.2.0). FreeSurfer was run in standard cross-sectional mode; the longitudinal processing stream was not applied. Raw DICOM data were converted to BIDS format using dcm2bids.
T1-weighted images underwent intensity nonuniformity correction, skull stripping, tissue segmentation, nonlinear registration to MNI space, and five-tissue-type segmentation. FreeSurfer-based surface reconstruction generated cortical segmentations, with subcortical, cerebellar, and cortical parcellations aligned to native and standard spaces.
Diffusion MRI data were preprocessed using MRtrix3 (dwifslpreproc) with FSL eddy for eddy-current and motion correction. Because the acquisition included a single b0 image without reverse phase-encoding, FSL TOPUP susceptibility correction was not applicable (-rpe_none option). Bias field correction was applied, and diffusion tensors were estimated to derive FA and MD maps. Fiber orientation distributions were calculated, followed by intensity normalization, structural-diffusion co-registration, and anatomically constrained tractography mask generation.
Quality control was performed by visual inspection of b0 images, FA maps, and track-density images for artifacts, motion, and tractography failures. Seven participants were excluded from an initial sample of 102, yielding a final dataset of 95 participants.
Structural connectomes were generated using probabilistic tractography (40 M streamlines, iFOD2 algorithm) with SIFT2 filtering. A 150-node atlas was used, comprising 100 cortical regions (Schaefer atlas), 34 cerebellar regions (SUIT atlas), and 16 subcortical regions (FreeSurfer). For each scan, a 150 × 150 connectivity matrix (11,175 undirected edges) was generated. Consistency-based thresholding retained edges present in ≥ 90% of all observations; edges absent in ≥ 10% were excluded prior to analysis.
Glymphatic-related measurements
Perivascular diffusivity was assessed using the DTI-ALPS method42. Directional diffusivity maps (Dxx, Dyy, Dzz) were generated using FSL’s dtifit (–save-tensor option). ALPS indices were computed using the automated DTI-ALPS pipeline (https://github.com/evavanheese/DTI-ALPS), which places bilateral 3 × 3 pixel ROIs at the level of the lateral ventricular body within projection and association fiber regions, minimizing operator-dependent variability.
Perivascular space (PVS) volumes were quantified using the PINGU framework43, a deep learning implementation of nnU-Net (v1.7.0). Binary PVS masks were generated, volumes calculated in mm³, and normalized by intracranial volume ((PVS volume in mm³/ICV in mm³) × 1,000).
CP volumes were derived from FreeSurfer automated segmentation and refined using a Gaussian Mixture Model segmentation approach (https://github.com/EhsanTadayon/choroid-plexus-segmentation)44. Normalized CP volume was calculated as: Normalized CP volume = (CP volume in mm³/eTIV in mm³) × 1,000.
NBR analysis
Longitudinal associations between glymphatic markers (ALPS index), clinical variables, and structural connectivity were examined using Network-Based R-statistics (NBR package, RRID: SCR_019114) in R, with linear mixed-effects models via the nlme package. Three complementary models were fitted:
Full cohort: Group × Time_Relative + Sex + Age + ALPS (random intercept).
PD only: Time_Relative + Sex + Age + ALPS + Disease_Duration + LEDD + UPDRSIII_OFF (random intercept ± slope).
Control only: Time_Relative + Sex + Age + ALPS (random intercept).
All continuous predictors were z-scored. Supra-threshold connections were clustered into connected components, and FWE correction was performed using 1,000 within-subject permutations (significance: FWE-corrected p < 0.05). Choroid plexus volume was excluded from final models.
LGC modeling
Edges showing significant Time_Relative effects in the PD-only NBR model were further examined using latent growth curve modeling (lavaan v.0.6–12). Linear models with latent intercept and slope factors were specified (intercept loadings = [1,1,1]; slope loadings = [0,1,2]), with Sex, Age, ALPS, Disease_Duration, UPDRSIII-OFF, and LEDD as covariates predicting both latent factors. A hierarchical strategy (full model → base model without covariates → constrained model with slope variance fixed to zero) was applied to ensure convergence; the first converging model with non-negative residual variances was retained. Models were estimated using ML with FIML for missing data. Slope significance was evaluated at α = 0.01, and results are reported as slope estimates with 95% confidence intervals.
Structural connectivity association analysis with motor severity
The relationship between structural connectivity and UPDRS-III motor scores was examined using 150 × 150 structural connectivity matrices, with two medial wall parcels excluded (148 ROIs). Regional node degree was calculated and z-scored within each medication state. Spearman rank correlations were computed separately for ON-state (n = 186) and OFF-state (n = 154) assessments, with FDR correction applied using the Benjamini–Hochberg procedure (q < 0.05). Regions surviving FDR correction were entered into multivariate Bayesian regularized linear regression models (brms package) with regularized horseshoe priors, 4 chains, 2,000 iterations (1,000 warmup), and adapt_delta = 0.99. Age was included as a covariate in all regression models. Convergence was assessed using R-hat (< 1.01) and effective sample size; model fit was evaluated using Bayesian R² with 95% credible intervals and LOO-ELPD. Posterior summaries were extracted using the bayestestR package. Sensitivity analyses adjusting for LEDD and disease duration yielded consistent results.
Statistical analysis
All analyses were performed in R. Statistical significance was set at p < 0.05 unless otherwise specified. Longitudinal changes in glymphatic and clinical measures were evaluated using linear mixed-effects models (nlme package) with Session, Age, and Sex as fixed effects, and participant ID as a random intercept. Model fit was assessed using AIC, BIC, and log-likelihood. Effect sizes are reported as η² for mixed-effects models and Bayesian R² (with 95% credible intervals) for Bayesian regression models. All analysis code is publicly available at GitHub.