Data used in the preparation of this article were obtained on [2025-3-25] from the Parkinson’s Progression Markers Initiative (PPMI) database,¹ RRID: SCR_006431. For up-to-date information on the study, visit http://www.ppmi-info.org. The PPMI cohort is a global observational study that longitudinally assesses participants recruited through neurology clinics associated with academic medical institutions across various countries. For our analysis, we included patients with a diagnosis of PD (Figure 1). The initial diagnosis was conducted by movement disorder specialists and on-site investigators and was subsequently reviewed and confirmed by a central consensus committee to ensure diagnostic accuracy. This study included a long-term follow-up of PD patients, with the longest follow-up period reaching up to 14 years.

PPMI studies were conducted in accordance with relevant regulations and guidelines. Written informed consent was obtained from all participants, and written informed consent was given to the subjects. The PPMI was approved by the local Institutional Review Boards of all participating agencies. All research procedures were conducted in accordance with the Declaration of Helsinki of 1964 and its subsequent amendments.

To investigate the longitudinal trajectories of motor symptoms in PD patients, we utilized the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale Part III (MDS-UPDRS Part III) scores, which comprehensively assess the severity of motor symptoms. The MDS-UPDRS Part III scores were obtained from the PPMI database, which includes detailed motor assessments conducted at regular intervals. These scores provide a quantitative measure of motor symptom severity, covering aspects such as tremor, rigidity, bradykinesia, and postural instability. The scores were collected during “OFF” medication states to capture the full spectrum of motor symptom variability.

The mean striatum dopamine transporter binding measure (DaTscan striatum) was calculated using DaTscan imaging data, which assesses dopamine transporter density in the striatum. Specifically, the mean striatum measure was derived by averaging the dopamine transporter binding values from the right and left caudate nuclei and putamen, as provided in the PPMI database.

The mild cognitive impairment (MCI) test scores were derived from a composite of cognitive tests administered to assess MCI in PD patients. The indicator for MCI was based on cognitive test scores, where a participant was classified as having MCI if at least two cognitive test scores were more than 1.5 standard deviations (SD) below the standardized mean. The specific cognitive tests included in the assessment were HVLT Total Recall, HVLT Recognition Discrimination, Benton Judgment of Line Orientation, Letter Number Sequencing, Semantic Fluency Test, and Symbol Digit Modalities. Participants were classified as having normal cognition if none of the cognitive test scores exceeded 1.5 SD below the standardized mean, and as having test-based MCI if at least two cognitive test scores exceeded this threshold.

The trajectories of MDS-UPDRS III score were analyzed using the latent class trajectory analysis to identify patterns of change over the follow-up period (Nagin et al., 2018). The most appropriate trajectory model was determined through a two-stage model selection process, taking into account factors such as trajectory shape (e.g., linear, quadratic, or cubic) and the number of groups. The number of dimensions and groups were optimized using Bayesian Information Criterion (BIC) values and clinical interpretability. Several models with varying numbers of trajectories (ranging from two to five) were attempted until the best-fitting model was identified. The performance of the candidate models is detailed in Supplementary Table 1. Once the best-fitting model was determined, participants were assigned to MDS-UPDRS Part III score trajectories based on the highest estimated probability of belonging to each group. The discrimination level of the average posterior probability was set at ≥ 0.70. The average posterior probability of participants assigned to each group was sufficiently high (ranging from 0.73 to 0.77), indicating a robust classification. Ultimately, three distinct trajectories were recognized.

The clinical characteristics of participants were summarized, stratified by the three trajectory groups identified based on the MDS-UPDRS III score. To scrutinize the dispersion of continuous variables, we deployed histogram analysis, Q-Q plots, and invoked the Kolmogorov–Smirnov test. Continuous variables adhering to a normal distribution were characterized by their mean ± SD, whereas those exhibiting skewness were depicted by the median (interquartile range, IQR). Categorical variables were elucidated through frequencies and percentages. When comparing continuous variables across the three trajectory clusters, we applied one-way analysis of variance (ANOVA) for normally distributed data and the Kruskal-Wallis H test for data with skewness. For categorical variables, we utilized the chi-square test or Fisher’s exact test, contingent on the data’s suitability, to discern differences among the groups.

To investigate the associations between various factors and the trajectories of MDS-UPDRS III score in PD patients, we conducted a series of statistical analyses after multiply imputing missing covariate values. For each factor of interest, including age, sex, body mass index (BMI), DaTscan striatum, MCI, serum glucose levels, albumin, age at onset of symptoms, and baseline MDS-UPDRS III score, we performed separate multinomial logistic regression analyses (Upton, 1995) to assess its independent association with the MDS-UPDRS III score trajectories. For each candidate predictor we fitted a separate multinomial-logistic regression in which that predictor was the exposure of interest and the remaining eight variables were simultaneously entered as covariates.

To assess the robustness of our findings, we conducted sensitivity analyses using two approaches. First, we excluded all records with missing data for the covariates and re-performed the correlation analysis to verify the stability of the results. Second, to further validate the results over a different time span, we utilized data from the first 10 years of follow-up to re-run the trajectory analysis and correlation analysis.

To investigate the role of various clinical indices in predicting the trajectories of MDS-UPDRS III scores in PD patients, we employed machine learning algorithms (Supplementary Figure 1). Prior to predictive modelling, we consolidated the three previously identified MDS-UPDRS III trajectories into a binary outcome to facilitate algorithm implementation and clinical interpretation. Group 1 (slow progression) was retained as the “slow” class, whereas Group 2 (moderate progression) and 3 (rapid progression) were merged to form a single “accelerated progression” class. The dataset comprised features such as age, sex, BMI, DaTscan striatum, MCI status, serum glucose levels, albumin, age at onset of symptoms, baseline MDS-UPDRS III score, gene-subtype of PD, 15-item Geriatric Depression Scale (GDS), Epworth Sleepiness Scale (ESS), and Rapid Eye Movement Sleep Behavior Disorder Questionnaire (RBD), and State–Trait Anxiety Inventory (STAI) total score. Initially, we utilized the Boruta algorithm for feature selection to identify key variables from this pool. Following the selection of significant features by Boruta, we constructed predictive models using these variables. Sporadic missing values in the feature matrix were handled by a training-set–specific simple imputation strategy: continuous variables were replaced with the training-fold mean, and categorical variables with the training-fold mode; these statistics were recalculated within each cross-validation split to prevent data leakage. To ensure robustness, models were trained with a phased integration framework involving human-provided data followed by machine processing for adaptive model architecture. We performed 5-fold cross-validation and utilized grid search for hyperparameter optimization. The optimal model was chosen based on the maximum Area Under the Receiver Operating Characteristic Curve (AUC). Random Forest (Wright and Ziegler, 2017; Probst et al., 2019) emerged as the best-performing model, which was further subjected to hyperparameter tuning and 5-fold cross-validation to ensure optimal performance and reliability. The effectiveness of the Random Forest model was assessed by comparing AUC values and SHAP beeswarm plots to evaluate the contribution of each feature to the model’s predictions.

All statistical analyses were conducted using R version 4.2.2 and the Free Statistics platform version 2.1 (Beijing Free Clinical Medical Technology Co., Ltd.), with results considered significant at a two-sided p value < 0.05, ensuring the transparency and reproducibility of our findings.

The results of latent class trajectory model construction are presented in Supplementary Table 1. It presented the results of BIC testing, posterior probabilities and classification proportion. For the five evaluated models, only the three-class model exhibited the best fit with the lowest BIC meanwhile the posterior probabilities were above 0.70 and the sample size in each class was above 5%. Thus, the three-class model was chosen as an optimal model. Figure 2 presents the three distinct trajectories of the longitudinal changes in MDS-UPDRS III score. Trajectories 1 to 3 account for 38% (461), 55.9% (679), and 6.1% (74) of the total sample, respectively. The variation of the three trajectories can be interpreted as: Trajectory 1, slow progression; Trajectory 2, moderate progression; Trajectory 3, rapid progression.

A total of 1,214 PD patients were ultimately included in the study following the trajectory analysis (Table 1). The majority of the patients were White (93.5%) with a mean age of 62.8 ± 9.7 years. The study cohort consisted of early-stage PD patients. Specifically, the distribution according to the Hoehn and Yahr (H&Y) stage was as follows: 0.3% (n = 3) at stage 0, 33.9% (n = 397) at stage 1, 64.0% (n = 749) at stage 2, and 1.9% (n = 22) at stage 3.

To investigate the factors associated with the trajectories of MDS-UPDRS III score in PD patients, we conducted multinomial logistic regression analyses, with the results presented in Figure 3. The analysis revealed that baseline MDS-UPDRS III score, albumin, BMI (per 5-unit increase), and DaTscan striatum were significantly associated with the trajectories of disease progression. Every one-point higher baseline MDS-UPDRS III score increased the likelihood of belonging to the Moderate-progression group by 27% (OR = 1.27, 95% CI: 1.23–1.31; p < 0.001) and to the Rapid-progression group by 49% compared with the Slow-progression group, after adjustment for all covariates (OR = 1.49, 95% CI: 1.43–1.57; p < 0.001). Similarly, each 1 mmoL/L decrement in serum albumin raised the odds of moderate progression by 5% (OR = 0.95, 95% CI: 0.91–0.99; p = 0.014) and of rapid progression by 11% (OR = 0.89, 95% CI: 0.81–0.98; p = 0.016). Every 5 kg/m² increase in BMI was associated with 19% higher odds of moderate (OR = 1.19, 95% CI: 1.01–1.41; p = 0.042), whereas no significant association was observed between BMI and rapid progression. Finally, each 1-unit lower DaTscan striatum reduced the odds of moderate progression by 32% compared with the slow-progression group (OR = 0.68, 95% CI: 0.46–0.99, p = 0.044). These findings indicate that baseline BMI, MDS-UPDRS III score, DaTscan striatum, and albumin are significant predictors of the rate of motor symptom progression in PD patients.

Sensitivity analyses were conducted to assess the robustness of our findings. First, we excluded all records with missing data (slow progression 450 [37.0%], moderate progression 704 [57.8%], rapid progression 63 [5.2%]) and re-performed the correlation analysis (Supplementary Figure 2). Second, we utilized data from the first 10 years of follow-up to re-run the trajectory (slow progression 601 [42.7%], moderate progression 714 [50.7%], rapid progression 94 [6.7%]) and correlation analyses (Supplementary Figure 3). Both approaches yielded consistent results, confirming the robust associations between the identified predictors and disease progression trajectories.

Specifically, lower albumin levels were associated with an increased likelihood of being in the Moderate Progression group (missing data exclusion: OR = 0.95, 95% CI: 0.91–0.99, p = 0.025; 10-year follow-up: OR = 0.95, 95% CI: 0.91–0.99, p = 0.016) and a significant trend towards decreased likelihood in the Rapid Progression group (missing data exclusion: OR = 0.87, 95% CI: 0.78–0.96, p = 0.007; 10-year follow-up: OR = 0.9, 95% CI: 0.82–0.98, p = 0.022) compared to the Slow Progression group. Higher baseline MDS-UPDRS III scores were strongly associated with both the Moderate Progression group (missing data exclusion: OR = 1.27, 95% CI: 1.23–1.31, p < 0.001; 10-year follow-up: OR = 1.26, 95% CI: 1.22–1.29, p < 0.001) and the Rapid Progression group (missing data exclusion: OR = 1.49, 95% CI: 1.42–1.57, p < 0.001; 10-year follow-up: OR = 1.47, 95% CI: 1.41–1.54, p < 0.001) compared to the Slow Progression group. DaTscan striatum was associated with an increased likelihood of being in the Moderate Progression group (missing data exclusion: OR = 0.95, 95% CI: 0.91–0.99, p = 0.025; 10-year follow-up: OR = 0.95, 95% CI: 0.91–0.99, p = 0.016) and a significant trend towards decreased likelihood in the Rapid Progression group (missing data exclusion: OR = 0.64, 95% CI: 0.42–0.98, p = 0.038; 10-year follow-up: OR = 0.65, 95% CI: 0.44–0.94, p = 0.022) compared to the Slow Progression group.

For BMI, the association with the Rapid Progression group was inconsistent, showing no significance with missing data exclusion (OR = 1.31, 95% CI: 0.94–1.8, p = 0.106), but borderline significance in the 10-year follow-up (OR = 1.33, 95% CI: 1–1.76, p = 0.046), whereas the association with the Moderate Progression group remained robust (missing data exclusion: OR = 1.22, 95% CI: 1.01–1.47, p = 0.038; 10-year follow-up: OR = 1.21, 95% CI: 1.02–1.43, p = 0.028).

During the feature selection phase, we utilized the Boruta algorithm to screen all covariates. As illustrated in Supplementary Figure 4, the variables located in the green area, namely age, BMI, mean striatum measure, serum glucose levels, albumin, age at onset of symptoms, and baseline MDS-UPDRS III score, were identified as key factors for the model. These 6 feature variables were ultimately employed to construct the model. When selecting the most effective machine learning method, AUC serves as a pivotal criterion. Consequently, the Random Forest model was deemed optimal. The Random Forest model exhibited an precision of 0.9285, an accuracy of 0.8587, and an F1 score of 0.8806. The AUC value for the training set was 0.950 and test set was 0.873, as shown in Figure 4. It is worth noting that the SHAP-beeswarm plot highlighted the significant roles of baseline MDS-UPDRS III score, albumin, BMI, and mean striatum volume, within the model (Figure 5).