Interpretable Machine Learning for Cross‐Cohort Prediction of Motor Fluctuations in Parkinson's Disease

Abstract

Background

Motor fluctuations are a common complication in later stages of Parkinson's disease (PD) and significantly affect patients' quality of life. Robustly identifying risk and protective factors for this complication across distinct cohorts could lead to improved disease management.

Objectives

The goal was to identify key prognostic factors for motor fluctuations in PD by using machine learning and exploring their associations in the context of the prior literature.

Methods

We applied interpretable machine learning techniques for time‐to‐event analysis and prediction of motor fluctuations within 4 years in three longitudinal PD cohorts. Prognostic models were cross‐validated to identify robust predictors, and the performance, stability, calibration, and utility for clinical decision‐making were assessed.

Results

Cross‐validation analyses suggest the effectiveness of the models in identifying significant baseline predictors. Movement Disorder Society‐Unified Parkinson's Disease Rating Scale parts I and II, freezing of gait, axial symptoms, rigidity, and pathogenic GBA and LRRK2 variants were positively correlated with motor fluctuations. Conversely, motor fluctuations were inversely associated with tremors and late age of onset of PD. Cross‐cohort data integration provides more stable predictions, reducing cohort‐specific bias and enhancing robustness. Decision curve and calibration analysis confirms the models' practical utility and alignment of predictions with observed outcomes.

Conclusions

Interpretable machine learning models can effectively predict motor fluctuations in PD from baseline clinical data. Cross‐cohort data integration increases the stability of selected predictors. Calibration and decision curve analyses confirm the model's reliability and utility for practical clinical applications. © 2025 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.

Motor fluctuations (MF) are a common complication in Parkinson's disease (PD), characterized by alternating periods of worsening and improving motor function. ¹ They have a significant negative effect on emotional well‐being and quality of life ² and are associated with increased healthcare costs ³ and higher hospitalization rates. ⁴ Management strategies include deploying complementary medications, dietary modifications, ⁵ and invasive therapies, such as deep brain stimulation (DBS) or pump systems. ⁶ , ⁷ The timing of MF varies considerably among patients, ⁸ with approximately half of them developing MF within 5 years. ⁹

Significant risk‐associated factors for developing MF include younger age at onset (AAO) ⁵ , ⁹ and the presence of rigidity and bradykinesia. ¹⁰ PD‐associated genetic variants, for example, in the GBA gene, ¹¹ , ¹² , ¹³ and gastrointestinal (GI) disorders have also been linked with MF. ⁵ , ⁸ , ¹⁴ Conversely, the tremor‐predominant subtype is associated with a decreased risk of MF. ⁹

Early and reliable prognosis of MF could pave the way for new therapeutic approaches to alleviate the progression of motor symptoms. ¹⁵ However, the disease‐modifying benefit of potential early medical, physical, dietary, or lifestyle adjustments or interventions still needs to be demonstrated. ¹⁴ , ¹⁶ For this purpose, identifying prognostic factors associated with MF could facilitate precision medicine trials to test preventive or early disease management strategies for improving patient outcomes. ¹⁷ , ¹⁸

Previous studies on MF have focused on individual cohorts, which can be prone to cohort‐specific biases. ¹⁹ To address this limitation, we analyze MF by comparing single‐ and multi‐cohort machine learning (ML) analyses. Cross‐cohort studies offer a more reliable identification of risk factors and exploration of their interactions by accounting for variability across different patient populations. By combining multi‐cohort analyses with interpretable ML approaches, we aim to provide more robust and insightful prognostic models that can pave the way for designing earlier and more effective therapeutic interventions.

Patients and Methods

Study Population

Three PD cohorts were covered in our analyses: the Luxembourg Parkinson's Study (LuxPARK, 395 participants), ¹⁹ the Parkinson's Progression Markers Initiative (PPMI, 485 participants), ²⁰ and the French ICEBERG cohort study (ICEBERG, 116 participants) ²¹ (see Supplementary Section S1 for inclusion/exclusion criteria). We included PD patients with or without MF, where MF was defined according to the Movement Disorder Society‐Unified Parkinson's Disease Rating Scale (MDS‐UPDRS) part IV. ²² , ²³

Furthermore, participants were required to meet the following inclusion criteria:

(1) A diagnosis of PD according to the United Kingdom Parkinson's Disease Society Brain Bank Diagnostic Criteria (UKPDSBB) criteria, ²⁴ or the presence of at least two of the following symptoms: resting tremor, bradykinesia, or rigidity, with either resting tremor or bradykinesia being one of the essential symptoms; or a single asymmetric resting tremor or asymmetric bradykinesia. ²⁵

(2) The presence of MF within 4 years following the baseline visit or the completion of a MF assessment without presenting symptoms within 4 years.

Participants were classified into MF+ (developed MF within 4 years from baseline) and MF− groups (no MF symptoms within the 4‐year). Specifically, inclusion in the MF+ group required a score of ≥1 on items 4.3 (time spent in off state), 4.4 (functional impact of fluctuations), or 4.5 (complexity of motor fluctuations). This information, available either from the stored MDS‐UPDRS assessments or from a corresponding dedicated clinical assessment variable capturing the same information, was consistently recorded in binary format (yes/no) for all participants.

Supplementary Table S1 details the number of patients per group meeting these criteria. Individuals with baseline MF were excluded from the validation set in the cross‐validation (CV) process and the external hold‐out set to prevent bias and overly optimistic predictive performance. Demographic and baseline clinical characteristics for the cross‐cohort analysis are shown in Supplementary Table S2, with cohort‐specific statistics for LuxPARK, PPMI, and ICEBERG in Supplementary Tables S3–S5.

ML Analysis of Motor Fluctuations

We applied ML approaches for classification and time‐to‐event analysis for MF prognosis in PD (see study workflow shown in Supplementary Fig. S1). The classification approaches predict MF occurrence within 4 years, whereas the time‐to‐event analysis estimates the duration until MF onset or the last follow‐up for censored data. Data preprocessing involved testing multiple cross‐study normalization methods (mean‐centering, ²⁶ standardization, quantile normalization, ²⁷ , ²⁸ ComBat, ²⁹ , ³⁰ Ratio‐A, ³¹ M‐ComBat, ³² see details in the Supplementary Section S3). The CV workflow, including model evaluation and data processing steps, is detailed in Supplementary Figures S2–S3. Variable aggregation was applied (see Supplementary Table S6) to combine related features that reflect similar underlying clinical features, thereby improving robustness and ensuring a more informative representation of predictors across cohorts. The total MDS‐UPDRS part IV score was excluded from the input to avoid circularity. Furthermore, because the LuxPARK cohort was assessed exclusively during the on state, only on state measurements were included.

Nine classification algorithms, focusing on decision tree/rule‐based methods were evaluated, including AdaBoost, ³³ , ³⁴ CART, ³⁵ CatBoost, ³⁶ C4.5, ³⁷ Fast Interpretable Greedy‐Tree Sums (FIGS), ³⁸ GOSDT‐GUESSES, ³⁹ GBoost, ⁴⁰ Hierarchical Shrinkage (HS), ⁴¹ and XGBoost. ⁴² For time‐to‐event analysis, we used component‐wise GBoost (CW‐GBoost), ⁴³ Survival Trees, ⁴⁴ Extra Survival Trees (EST), ⁴⁵ Survival GBoost, ⁴⁶ LSVM, NLSVM, ⁴⁷ Penalized Cox, ⁴⁸ , ⁴⁹ and Survival Random Forests (SRF). ⁵⁰ Classification models incorporated the scikit‐learn (v1.2.2), imodels (v1.2.6), CatBoost (v1.1), and XGBoost (v1.6.2) packages, whereas scikit‐survival (v0.20.0) was used for time‐to‐event models. All preprocessing methods, including imputation (multiple imputation by chained equations [MICE] method for variables with <50% missingness) and cross‐study normalization, were performed independently in the training and validation sets to avoid data leakage (Supplementary Fig. S3). In each outer fold, an inner fold handled hyperparameter tuning and feature selection for model optimization (see Fig. 4 and Supplementary Section S4).

FIG. 1.

Illustration of the machine learning and cross‐validation workflow. The workflow involves training and evaluating machine learning models for motor fluctuation prognosis using 5‐fold cross‐validation to assess average performance. A 3‐fold nested cross‐validation on the training set data was used to optimize hyperparameters and select the most informative features. [Color figure can be viewed at wileyonlinelibrary.com]

We observed varying degrees of imbalance in MF outcomes across our study cohort (see Results for details). To address this, we implemented an undersampling technique during model training. This involved randomly removing samples from the majority class (MF+) to achieve a more balanced representation of outcomes. Although this reduces the overall sample size, it improves the model's ability to learn patterns associated with the minority class (MF−).

To ensure cross‐cohort comparability, MF analyses used only shared baseline features across all cohorts (see the list of features in Supplementary Tables S2–S5). Two types of models were trained independently:

(1) Comprehensive model, which incorporates all shared baseline features across cohorts, capturing potential interactions and nonlinearities influencing MF.

(2) Refined model, which excludes baseline MF and levodopa intake to facilitate discovery of new predictors and improve model interpretability and robustness.

Furthermore, we refer to the best model as the one achieving the highest average area under the receiver‐operating characteristic (ROC) curve (AUC) (classification), or concordance index (C‐index) (time‐to‐event analysis) across the hyperparameter optimization within the nested CV.

Model Interpretation

Shapley additive explanations (SHAP) value analysis ⁵¹ was used to interpret the MF prediction models by quantifying each variable's predictive impact. Furthermore, we examined the median conversion time for 50% of PD patients to develop MF, using log‐rank tests to compare Kaplan–Meier (KM) curves between groups for significant differences in time‐to‐MF occurrence.

Hazard ratios (HR) were derived from SHAP scores to quantify the relative MF risk, highlighting key associations. Confidence intervals (CI) for HR estimates were calculated using bootstrapping and the percentile method at a 5% significance level. ⁵² Additionally, the log‐rank test was used to determine the optimal threshold for continuous variables, enabling HR calculation by stratifying patients into high‐ and low‐risk groups.

Predictive Model Evaluation

MF prediction models were evaluated using the AUC for classification performance and the C‐index for time‐to‐MF analysis. The AUC (ranging from 0–1, where 0.5 indicates no discriminative ability, values below 0.5 suggest worse than random performance, and 1.0 indicates perfect discrimination) was used a measure of a model's discriminative ability in classification tasks, whereas the C‐index (with a similar range and interpretation) served as a comparable measure of discrimination in time‐to‐event models. Both metrics offer advantages over simpler performance measures such as accuracy, because they are insensitive to class imbalance and provide a more comprehensive assessment of model performance across different classification thresholds. DeLong's test and one‐shot nonparametric tests were used to compare hold‐out test set AUCs and C‐indices to identify significant differences. Using these statistical tests, we evaluated the impact of excluding the variables “levodopa medication” and “baseline MF” and compared models with and without cross‐study normalization (normalized vs. unnormalized model). P‐values were adjusted for multiple comparisons. ⁵³ Furthermore, Bayesian signed‐rank tests were applied to compare the best models' cross‐validated performances across cohorts, ⁵⁴ whereas model stability was assessed by examining the standard deviation (SD) across CV cycles.

Analysis of Feature Importance across Multiple Cohorts

We analyzed feature selection statistics for MF classification and time‐to‐MF analysis. First, the feature selection frequency was determined across CV cycles for each cohort. We, then, compared mean selection percentages across LuxPARK, PPMI, and ICEBERG cohorts to identify the most consistent predictors.

Statistical Analyses

Hypothesis tests were used to identify between‐group differences in baseline clinical characteristics: the Mann–Whitney U test for non‐normally distributed variables, the two‐sample t test for normally distributed variables, and Fisher's exact test for categorical variables. Normality was assessed using the Shapiro–Wilk test. Analysis of variance (ANOVA) combined with Tukey's honestly significant difference test was applied to compare normally distributed data across cohorts, whereas the Kruskal‐Wallis test combined with Dunn's test was used for non‐normal data.

Correlation analysis used Spearman's correlation for continuous/ordinal variables, point‐biserial correlation for binary and continuous/ordinal variables, Matthew's correlation coefficient (MCC) for binary variables, and Kendall's τ for ordinal variables. Statistical significance was defined as P < 0.05.

Decision Curve and Calibration Analysis

Decision curve analysis (DCA) on the hold‐out test set was used to evaluate the clinical utility of predictive models by comparing the net benefit across decision thresholds when using the model to make treatment decisions versus treating all patients or none, ⁵⁵ , ⁵⁶ assuming adequate interventions are available.

The area under the net benefit curve (AUNBC) quantifies the model's clinical utility, with a larger AUNBC indicating a greater decision‐making advantage. AUNBC differences were tested using bootstrapped hypothesis testing with 1000 replicates for P‐value estimation. ⁵⁷

Additionally, calibration analysis was performed to assess the agreement between predicted probabilities and observed outcomes for MF classification. ⁵⁶ For time‐to‐MF outcomes, predicted conversion probabilities were compared to KM estimates at year 4. ⁵⁸ Calibration slope and mean squared error (MSE) were examined to ensure accuracy and reliability.

Results

Single‐Cohort Analyses

Among the three cohorts, the PPMI cohort provided the highest classification performance, with the best comprehensive MF model reaching a cross‐validated AUC of 0.71 ± 0.04 and a hold‐out AUC of 0.71 (Supplementary Table S7). The refined model (excluding baseline MF and levodopa medication) yielded a cross‐validated AUC of 0.70 ± 0.05 with a hold‐out AUC of 0.63 (Supplementary Table S8). For time‐to‐MF analysis, the comprehensive model achieved cross‐validated and hold‐out C‐indices of 0.71 ± 0.05 and 0.72 (Table 1), respectively, whereas the refined model had C‐indices of 0.69 ± 0.05 and 0.70 (Supplementary Table S9).

TABLE 1.

Summary of predictive performance metrics for comprehensive time‐to‐motor fluctuations

Single‐cohort analyses
Algorithm	LuxPARK			PPMI			ICEBERG
	Mean (SD)	Hold‐out C‐index	No. of features	Mean (SD)	Hold‐out C‐index	No. of features	Mean (SD)	Hold‐out C‐index	No. of features
CW‐GBoost	0.669 (0.134)	0.606	9 (9)	0.708 (0.022)	0.705	11 (23)	0.587 (0.094)	0.594	5 (13)
Extra Survival	0.587 (0.139)	0.629	10 (10)	0.711 (0.047)	0.718	108 (114)	0.575 (0.145)	0.642	10 (10)
Survival GBoost	0.598 (0.081)	0.588	81 (88)	0.681 (0.044)	0.684	12 (17)	0.628 (0.093)	0.687	12 (19)
LSVM	0.605 (0.161)	0.608	8 (8)	0.679 (0.060)	0.726	27 (27)	0.482 (0.068)	0.600	11 (11)
NLSVM	0.532 (0.173)	0.601	16 (16)	0.695 (0.043)	0.736	30 (30)	0.501 (0.105)	0.661	15 (15)
Penalized Cox	0.533 (0.144)	0.565	1 (2)	0.700 (0.036)	0.710	29 (37)	0.518 (0.049)	0.624	14 (14)
Survival RF	0.569 (0.109)	0.626	9 (9)	0.687 (0.013)	0.716	65 (94)	0.596 (0.046)	0.603	11 (11)
Survival Trees	0.555 (0.050)	0.595	14 (20)	0.640 (0.032)	0.637	6 (7)	0.538 (0.054)	0.633	6 (8)

Multi‐cohort analyses
Algorithm	Cross‐Cohort			Leave‐ICEBERG‐out			Leave‐PPMI‐out			Leave‐LuxPARK‐out
	Mean (SD)	Hold‐out C‐index	No. of features	Mean (SD)	Hold‐out C‐index	No. of features	Mean (SD)	Hold‐out C‐index	No. of features	Mean (SD)	Hold‐out C‐index	No. of features
CW‐GBoost	0.646 (0.035)	0.649	8 (9)	0.687 (0.025)	0.667	16 (27)	0.630 (0.065)	0.698	13 (24)	0.734 (0.032)	0.625	15 (24)
Extra Survival	0.640 (0.045)	0.680	158 (160)	0.685 (0.011)	0.588	144 (150)	0.644 (0.033)	0.698	21 (21)	0.729 (0.032)	0.617	123 (123)
Survival GBoost	0.640 (0.032)	0.654	18 (29)	0.660 (0.069)	0.656	11 (24)	0.616 (0.043)	0.690	12 (21)	0.712 (0.024)	0.560	29 (57)
LSVM	0.627 (0.038)	0.687	42 (42)	0.674 (0.033)	0.589	57 (57)	0.632 (0.036)	0.696	40 (40)	0.725 (0.039)	0.572	53 (53)
NLSVM	0.614 (0.033)	0.648	47 (47)	0.672 (0.043)	0.612	55 (55)	0.625 (0.045)	0.661	32 (32)	0.727 (0.039)	0.586	54 (54)
Penalized Cox	0.589 (0.082)	0.614	2 (3)	0.671 (0.046)	0.500	1 (3)	0.607 (0.042)	0.590	2 (4)	0.749 (0.037)	0.585	83 (118)
Survival RF	0.641 (0.034)	0.664	90 (129)	0.681 (0.037)	0.565	24 (55)	0.619 (0.030)	0.697	101 (115)	0.728 (0.034)	0.623	71 (103)
Survival Trees	0.590 (0.058)	0.589	2 (2)	0.629 (0.064)	0.489	4 (6)	0.597 (0.043)	0.642	10 (18)	0.665 (0.036)	0.614	8 (10)

The table presents cross‐validated and hold‐out C‐indices for single and multi‐cohort analyses and the number of features used in each model. Models with the highest average cross‐validated C‐indices in each cohort analysis across the nested cross‐validation (CV) hyperparameter optimizations are highlighted in bold and are referred to as the best models. The model with the highest hold‐out C‐index is indicated in italics. The “number of features” column includes the number of candidate features selected during CV (shown in brackets) and the subset of features demonstrating significant predictive impact identified through permutation importance analysis (preceding the brackets).

Abbreviations: LuxPARK, Luxembourg Parkinson's Study; PPMI, Parkinson's Progression Markers Initiative; ICEBERG, French ICEBERG cohort study; SD, standard deviation; No, number; CW‐GBoost, component‐wise Gradient Boosting; CV, cross‐validation; Survival GBoost, Survival Gradient Boosting; LSVM, Linear Support Vector Machine; NLSVM, Naive Linear Support Vector Machine; RF, Random Forest.

Lower predictive performances observed for LuxPARK and ICEBERG (Supplementary Figs. S4–S7) can be explained by the more imbalanced MF outcome distribution in LuxPARK (71.1% MF+) and a limited sample size in ICEBERG. In addition, differences in patient characteristics may have influenced the results, with participants from LuxPARK presenting with longer disease duration and higher MDS‐UPDRS scores than PPMI and ICEBERG (Supplementary Table S10), and PPMI presenting with the lower MDS‐UPDRS scores overall.

The cohort datasets showed varying but generally low levels of average missingness (LuxPARK: 2.1%, PPMI: 0.7%, and ICEBERG: 4.6%), suggesting a limited impact on the observed predictive performances.

Informative Predictors from Single‐Cohort Analyses

Key MF predictors across cohorts include disease duration, AAO, body weight, axial symptoms, MDS‐UPDRS parts I and II, Benton Judgment of Line Orientation (JLO), and autonomic dysfunction, including GI (Supplementary Table S11). Supplementary Table S12 highlights the consistently selected features across algorithms and cohorts, further emphasizing the importance of disease duration, MDS‐UPDRS parts I and II, and body weight as key predictors. These variables are interrelated, with disease duration influencing many covariates, including axial symptoms, MDS‐UPDRS scores, Scale for Outcomes in Parkinson's disease for Autonomic symptoms (SCOPA‐AUT), and disease severity features in general.

Multi‐Cohort Analyses

Multi‐cohort models showed higher robustness of performance statistics and generalizability than the single‐cohort models, despite slightly lower average predictive performance than PPMI‐only models. The cross‐cohort classification model with the highest performance reached a cross‐validated AUC of 0.64 ± 0.01 and a hold‐out AUC of 0.62 (Supplementary Table S7). For time‐to‐MF analysis, the corresponding models provided a cross‐validated C‐index of 0.65 ± 0.04 and a hold‐out C‐index of 0.65 (Table 1).

Despite variation in performance when single cohorts were left out, multi‐cohort models generally provided more stable results in the CV. Supplementary Table S13 presents the significance of differences between the performance statistics for different model types (comprehensive vs. refined). In addition, we investigated the effect of cross‐study normalization (mean‐centering), showing that it significantly improves hold‐out performance in the cross‐cohort time‐to‐MF models (P = 0.011) (Supplementary Tables S13–S14).

Overall, multi‐cohort analyses provided more stable and robust models than single‐cohort analyses, although the PPMI cohort reached the best performance. Despite slightly lower performance, multi‐cohort models may still be preferable because of their greater stability (Supplementary Figs. S8–S11) and generalizability across different patient populations.

Informative Predictors from Multi‐Cohort Analyses

To identify informative clinical features for MF prognosis, we analyzed their predictive impact and statistical associations. In the cross‐cohort analyses, levodopa medication, disease duration, the presence of dyskinesia, and resting tremor were the most frequently selected features (Supplementary Table S15). Levodopa medication, disease duration, and rigidity were consistently linked to higher MF likelihood, whereas resting tremor, and Benton JLO (higher scores indicate better visuospatial ability) were negatively correlated with MF severity (Fig. 2, Supplementary Fig. S12, and Supplementary Table S16 for the correlations with MF severity).

FIG. 2.

Shapley additive explanations (SHAP) value plot illustrating the top individual predictors for the best comprehensive model in cross‐cohort motor fluctuations classification. SHAP plot, providing a detailed view of how individual predictors influence the model's predictions. This visualization allows us to understand not just the overall importance of each predictor, but also how different values of that predictor impact the likelihood of motor fluctuations across our dataset. Each row represents a predictor, ordered by overall importance from top to bottom, and each point represents an individual observation in the dataset. The color of the points ranges from blue (lower values) to red (higher values) for that predictor. The position of points on the x‐axis shows the impact on the model's prediction: points to the right of zero indicate the predictor pushed the model toward predicting motor fluctuations, and points to the left of zero indicate the predictor pushed the model away from predicting motor fluctuations. Therefore, the SHAP plot shows whether high or low values of a predictor are associated with a higher or lower likelihood of predicted motor fluctuations. [Color figure can be viewed at wileyonlinelibrary.com]

In the time‐to‐MF analysis, key predictors included levodopa medication, disease duration, and dyskinesia (Fig. 3 and Supplementary Fig. S13). Although HR for these factors in the best comprehensive time‐to‐MF model did not reveal substantial differences between patient groups, log‐rank tests showed significant differences for disease duration and dyskinesia. Levodopa, although a key predictor, did not display significant differences in the time‐to‐MF analysis (Table 2). We note that disease duration significantly correlates with other features, including AAO, axial symptoms, bradykinesia, dyskinesia, the Hoehn and Yahr (H&Y) stage, and rigidity (Supplementary Table S17), highlighting the complexity of PD progression.

FIG. 3.

Shapley additive explanations (SHAP) value plot illustrating the most informative predictors for the best comprehensive model in the cross‐cohort time‐to‐motor fluctuation analysis. SHAP plot, providing a detailed view of how individual predictors influence the model's predictions. This visualization allows us to understand not just the overall importance of each predictor, but also how different values of that predictor impact the likelihood of motor fluctuations across our dataset. Each row represents a predictor, ordered by overall importance from top to bottom, and each point represents an individual observation in the dataset. The color of the points ranges from blue (lower values) to red (higher values) for that predictor. The position of points on the x‐axis shows the impact on the model's prediction: points to the right of zero indicate the predictor pushed the model toward predicting motor fluctuations, and points to the left of zero indicate the predictor pushed the model away from predicting motor fluctuations. Therefore, the SHAP plot shows whether high or low values of a predictor are associated with a higher or lower likelihood of predicted motor fluctuations. [Color figure can be viewed at wileyonlinelibrary.com]

TABLE 2.

A summary of hazard ratio (HR), median conversion times with 95% confidence interval (CI), and P‐values from the log‐rank test for the top 15 predictors in the cross‐cohort time‐to‐MF model is presented below.

Predictors	HR (95% CI)	Median conversion times (95% CI)	Log‐rank (P‐values)
Dyskinesia score
≥1	1.00 (0.8–2.37)	0.95 (0–2.44)	1.44E−06
<1		4.07 (3.77–4.36)
Levodopa treatment
Yes	1.14 (1.00–1.56)	3.01 (2.27–4.07)	0.446
No		4.19 (3.77–4.51)
MDS‐UPDRS I‐fatigue (normal)
Yes	0.70 (0.51–0.94)	4.44 (4.03–4.76)	0.116
No		3.03 (2.34–4.00)
MDS‐UPDRS IV‐painful off‐state dystonia (slight)
Yes	5.06 (1.72–16.23)	0 (0–0)	9.87E−08
No		4.01 (3.44–4.28)
Pathogenic LRRK2 variant
Yes	2.02 (1.41–2.92)	0.99 (0.11–2.03)	2.15E−06
No		4.11 (3.84–4.44)
Pathogenic GBA variant (pathogenic variant)
Yes	3.78 (2.13–6.26)	1.02 (0.28–2.26)	5.45E−07
No		4.11 (3.79–4.44)
Disease duration since PD diagnosis (years)
≥2	1.10 (0.98–1.49)	2.76 (2.03–4.01)	0.011
<2		4.28 (4–4.68)
MDS‐UPDRS I‐depressed moods (mild)
Yes	1.45 (1.00–2.57)	3.01 (0–5.02)	0.167
No		4.03 (3.31–4.28)

The HR indicates the risk associated with each predictor, whereas the median conversion time and the log‐rank test assess the Kaplan–Meier curve differences between groups.

Abbreviations: HR, hazard ratio; CI, confidence interval; MF, motor fluctuations; MDS‐UPDRS, Movement Disorder Society‐Unified Parkinson's Disease Rating Scale; PD, Parkinson's disease.

Additionally, fatigue and dystonia were associated with increased MF likelihood (HR of 1.4 and 5.1, respectively). Genetic factors were also significant predictors, including pathogenic LRRK2 variants (HR, 2.0; 95% CI, [1.4–2.9]), and pathogenic GBA variants (HR, 3.8; 95% CI, [2.1–6.3]).

Levodopa equivalent daily dose (LEDD) analysis showed that MF+ patients had a significantly higher mean baseline LEDD (780.7 mg) compared to MF− patients (587.2 mg) (P = 4.5E−04) (Supplementary Table S18). Moreover, time‐to‐MF was significantly longer for patients with LEDD <400 mg (2.3 years) compared to those with LEDD ≥400 mg (1.3 years, P = 3.2E−03).

Decision Curve and Calibration Analysis

DCA and calibration analysis were applied to assess model reliability and clinical utility. Both classification models, such as GBoost and AdaBoost, and time‐to‐MF analysis models, such as LSVM and NLSVM, were well calibrated (Supplementary Table S19) and showed higher net benefits than simple “treat all” or “treat none” strategies (Supplementary Figs. S14–S17), supporting the models' potential for individualized risk assessment.

Discussion

Predicting MF in PD is challenging because of interindividual variability and a multitude of influencing factors. Our study addresses these complexities by evaluating the robustness and accuracy of interpretable prediction models across multiple cohorts. Instead of focusing on a single approach, we assessed several multivariable ML classification and time‐to‐MF models. In addition, we ranked predictors robustly using feature selection statistics across CV cycles and cohorts. Below, we discuss the potential of these ML models as prognostic tools and interpret the identified predictors.

Comparative Evaluation of Predictive Models

Previous studies predicting MF in PD have achieved varying degrees of success. A few candidate prognostic factors have been highlighted, for example, one study reported disease duration and levodopa dose as key predictors. ⁵⁹ A further study confirmed that longer disease duration and younger AAO are significant predictors. ⁶⁰ However, most prior studies relied on a single cohort with small sample sizes and a large number of variables, which can lead to model overfitting, potentially affecting generalizability. To help address these limitations, we used multivariable ML approaches, cross‐cohort analyses, and conducted external evaluations.

We used various modeling approaches, each with distinct advantages. Tree‐based algorithms such as AdaBoost, CatBoost, and XGBoost handle complex variable interactions for classification well, while support vector machines (LSVM and NLSVM) perform well in time‐to‐event analyses because of their ability to handle censored information. Although we focused on modeling approaches providing interpretable information on feature relevance, future studies might still consider more complex approaches, such as meta‐learning combinations of multiple algorithms, to further improve predictive performance.

Given the variability in MF severity score distributions across cohorts, our approach focused on a binary classification approach instead of modeling MF severity as a continuous variable. Although severity scores may provide more specific results, the distinctive score distributions suggest that severity assessments may differ strongly between studies, potentially leading to inconsistencies. Conversely, binary classification facilitates cross‐cohort comparability and enables performance evaluation using standard ROC‐based metrics, providing both robust and comparable assessments of discriminative ability.

Our results highlight the value of cross‐cohort data integration, which improved model robustness and generalizability. Although cross‐study prediction is more challenging than building models for a single cohort, the best‐performing cross‐cohort model achieved AUCs/C‐indices around 0.7, indicating a significant, although likely still improvable, discriminative ability. Moreover, the comparatively higher performance achieved for the PPMI cohorts indicate the benefits of large sample sizes and extended follow‐ups. Despite the fact that our longitudinal cross‐study classification presents additional challenges compared to previous cross‐sectional, single‐study analyses, the predictive power for MF prognosis in our study is comparable to that observed in previous studies on motor complications in PD. For instance, previous studies reached an AUC of 0.68 for dyskinesia prognosis, ⁶¹ an accuracy of 72% for cross‐sectional MF detection, ⁵⁹ and AUC scores between 60% and 82% for cross‐sectional discrimination between on and off medication states. ⁶²

The predictive models for MF classification and time‐to‐MF analysis also demonstrate superior net benefit over simple strategies such as “treat all” or “treat none” in DCA, supporting their potential role in risk‐based patient management. By identifying patients at higher risk of MF, these models may aid in designing earlier intervention strategies, including pharmacological or lifestyle interventions, to delay or mitigate symptom progression. In addition, calibration analysis confirmed a strong agreement between predicted and observed outcomes, supporting the reliability of the models for real‐world applications. This agreement indicates the potential for optimized versions of these models to be integrated into clinical decision making processes to improve patient stratification for precision medicine trials of MF‐targeted interventions. Future research should explore the prospective validation of such optimized models in clinical settings to assess their utility in guiding treatment strategies.

However, important challenges remain for clinical application. Variability in performance across cohorts suggests that cohort‐specific biases influence model accuracy. Identified prognostic features such as LEDD, disease duration, and non‐motor symptoms underscore the complexity of predicting MF onset, because these are known to vary widely among patients. ⁶³ Although these models could inform future precision medicine studies, further optimization and external validation is still needed.

Interpretation of Models and Predictive Features

When interpreting baseline clinical features as MF prognostic predictors, it is important to consider the variations in baseline characteristics across the LuxPARK, PPMI, and ICEBERG cohorts. Longer disease durations in LuxPARK suggest current clinical characteristics in this cohort may have a greater influence on MF development. In contrast, lower average body weight and body mass index in ICEBERG may influence motor symptom progression because of potential differences in lifestyle, physical activity, and dietary habits. Although PPMI shares similar disease durations with ICEBERG, it displays significantly lower MDS‐UPDRS and SCOPA‐AUT scores, indicating other factors may be implicated in modulating symptom severity and non‐motor symptoms. Because of these cohort‐specific characteristics, our focus on cross‐cohort modeling ensures a more balanced interpretation of MF predictors.

Previous research has linked levodopa medication to MF, ⁶⁴ , ⁶⁵ but a recent trial contradicted these findings. ⁶⁶ Our univariate analysis showed a significant but potentially indirect association, with higher LEDD (≥400 mg) correlating with increased MF risk. ⁹ , ¹⁵ However, multivariate ML models did not find a statistically significant difference in MF risk on baseline levodopa use, and KM curves indicated no significant impact. These results emphasize the value of multivariate models in capturing complex interactions that univariate analyses may overlook.

Significant correlations between levodopa intake at baseline, disease duration, and H&Y stage of 0.47 and 0.34 (P = 5.4E−55 and 1.5E−27), respectively, ¹⁴ suggest that levodopa's association with MF may be indirectly linked to disease progression (we note that the variable levodopa refers only to levodopa intake at baseline, not to LEDD). Besides its role in MF risk, higher LEDD correlated with greater MF severity (Spearman correlation in LuxPARK of 0.42, P = 1.7E−18), further supporting its association with symptom severity. Furthermore, the comprehensive and refined models demonstrated comparable hold‐out performance in cross‐cohort analyses, with no statistically significant differences. This suggests that, despite the apparent influence of levodopa in univariate analyses, its predictive value may be overshadowed by other strongly correlated predictors. The performance of the refined model indicates that predictors associated with levodopa, such as disease duration, can be sufficient for maintaining predictive capability when levodopa is excluded.

Disease duration and H&Y stage were consistently associated with an increased risk of developing MF, ¹² reflecting associations between disease stage and motor impairments. ⁶⁷ In addition, longer disease duration was significantly correlated with greater MF severity, further emphasizing its association with overall symptom burden (Spearman correlation of 0.41, P = 1.0E−38). Moreover, an earlier AAO was associated with longer disease duration (Spearman correlation of −0.22, P = 3.4E−12), which in turn was associated with a higher risk of MF (point‐biserial correlation of 0.3, P = 6.7E−22; see beginning of article). ⁵ , ⁹ These results highlight the association of disease progression with both the onset and severity of MF and the relevance of disease duration and staging in identifying high‐risk patients for targeted management strategies.

Higher baseline MDS‐UPDRS part II scores, reflecting severe motor impairments such as freezing of gait (FoG), are associated with the development of MF. ⁶¹ This is expected, as FoG often occurs during off‐phases, which are central to MF. Impaired movement speed, gait instability, and axial symptoms, including gait impairments, and rigidity, derived as aggregated measures from MDS‐UPDRS part II and III assessments, have also been linked with MF onset and severity. ¹⁰ , ²² , ⁶⁸ Conversely, the observed negative correlation between tremors and MF may reflect the milder disease progression in tremor‐dominant patients. ⁶⁹ In addition, superior visuospatial ability (Benton JLO) was associated with a lower risk of MF, potentially because of its association with reduced risk of FoG and lower tremor severity. ⁷⁰ , ⁷¹

MDS‐UPDRS part I scores, capturing non‐motor symptoms, were also associated with an elevated MF risk, likely because of the high correlation with motor symptoms and general disease severity. ⁶¹ Furthermore, fatigue and sleep disturbances (including insomnia and excessive daytime sleepiness) were found significantly correlated ⁷² and to be positively associated with MF, ⁷³ possibly because of their association with overall disease burden and medication response. ⁷⁴ , ⁷⁵ Patients experiencing fatigue are more likely to develop apathy and depression, ⁷⁶ suggesting common underlying pathophysiological mechanisms, potentially linked to dopaminergic dysfunction. ⁷⁷ Fatigue symptoms have been associated with impaired dopaminergic neurotransmission in the striatum, ⁷⁸ a finding that may be affecting the levodopa response, ⁷⁹ and consequently, the risk of MF. ⁸⁰ Similarly, autonomic dysfunction showed positive MF associations, in line with previous studies. ⁸ , ⁸¹ Among autonomic symptoms, urinary dysfunction and GI correlate with the severity of motor and non‐motor impairments ⁸² , ⁸³ and have been proposed as early markers of motor and non‐motor disability in PD, ⁸⁴ , ⁸⁵ preceding the onset of more pronounced motor complications.

Finally, considering genetic factors, the cross‐cohort analysis revealed that pathogenic GBA mutations are associated with increased risk of MF, ¹³ potentially reflecting the more aggressive disease progression in GBA mutation carriers. ¹¹ , ¹³ Conversely, although pathogenic LRRK2 mutations are also associated with MF, ⁸⁶ the effect size is lower (HR, 2.0 and 3.8 for LRRK2 and GBA). GBA mutation carriers also show more severe motor progression compared to LRRK2 mutation carriers, ⁸⁷ a more advanced disease stage as measured by H&Y, ⁸⁸ , ⁸⁹ which is positively correlated with MF severity (Spearman correlation of 0.19, P = 3.3E−09). In addition, both GBA and LRRK2 mutations have been linked to dyskinesia, ⁹⁰ a common PD complication linked to MF. This highlights the multifaceted impact of these genetic variants, disease progression, and MF severity, and indicates the need to consider these factors when assessing motor complications in PD.

Although some of our findings, such as the association of LEDD, disease duration, and disease severity with MF risk, align with the clinical experience of many practitioners and previous research, our study goes beyond confirming these clinical observations in several ways. First, we provide precise quantifications of these associations, which can inform evidence‐based clinical decision‐making. Second, by demonstrating these patterns across multiple cohorts, we offer robust, generalizable evidence for these clinical observations. Third, our comprehensive approach has identified less obvious predictors, such as specific genetic variants, which may not be as readily apparent in routine clinical practice. Finally, by combining these factors into robust, multi‐factorial predictive models validated across multiple cohorts, we offer a tool that goes beyond simple association to provide individualized risk prediction. These models may, therefore, help to lay the ground for the design of precision medicine trials and prognostic applications to enable earlier and potentially more effective interventions against MF in PD.

Practical Applications in Clinical Trial Design

The predictive models developed in this study have several potential applications in improving protocol development and participant recruitment for clinical trials focused on MF in PD. First, our models can aid in risk stratification of potential participants for upcoming precision medicine trials. By identifying patients at higher risk of developing MF, researchers can enrich their study population with individuals more likely to experience the outcome of interest. This could potentially reduce the required sample size and duration of trials, making them more efficient and cost‐effective. Second, the identified predictors can inform inclusion and exclusion criteria for trials. For instance, trials testing interventions to prevent or delay MF might specifically target patients with MF‐associated characteristics derived from our models, for example, patients with PD‐associated GBA and LRRK2 variants, given their higher risk of developing MF. Last, these models could be used to develop personalized follow‐up schedules in trials. Participants predicted to be at higher risk of MF might be monitored more frequently, allowing for more timely detection of MF onset.

Summary and Conclusions

This study combined interpretable ML techniques and data from three cohorts to identify robust MF predictors in PD. Cross‐cohort data integration improved model generalizability, reduced cohort‐specific biases, and improved overall stability and robustness. Moreover, decision curve and calibration analyses showed that the models offer greater benefits for clinical decision‐making than simple “treat all” or “treat none” strategies.

Among the key predictors, baseline LEDD was strongly associated with MF, likely because of its association with disease duration and progression. Furthermore, disease severity and disease stage, as indicated by variables such as MDS‐UPDRS part I scores, H&Y stage and symptoms associated with late disease stage, such as dyskinesia, were important predictors, and subjects with tremor‐dominant PD showed a less severe progression and delayed onset of MF.

Overall, this study highlights the potential of interpretable ML models for predicting MF in PD, emphasizing the benefits of cross‐cohort data integration for improving model stability and generalizability. Despite promising results in terms of statistical significance and net benefit of the models, variability in hold‐out performance across cohorts remains a challenge. These predictive models offer practical tools for future clinical trial design in PD research. By helping to identify patients at higher risk of developing MF, they can contribute to more targeted and efficient trials of preventive or early interventions. However, it is important to note that these models are imperfect predictors, and their use should be combined with expert clinical judgment and consideration of other relevant factors in trial design and patient care. Follow‐up research should further optimize and validate the predictive models across more diverse cohorts to increase their value design of future precision clinical trials. Moreover, although a standardized methodology was used to assess MF, capturing general MF risk patterns and facilitating cross‐cohort comparisons, this does not differentiate between more specific symptoms, such as gait disturbance versus tremor fluctuations. Future prospective studies incorporating more precise symptom tracking, such as patient diaries or wearable sensors, could provide deeper insights into symptom‐specific fluctuations and at the same time reveal features that are more robust for ML applications. These tools could provide a clearer understanding of specific patterns of MF and their associations with underlying predictors.

In conclusion, our study provides a comprehensive, data driven analysis of predictors for MF in PD. Although some of our findings, such as the importance of LEDD, disease duration, and disease severity, align with well‐established clinical experience, our work provides precise quantification of these associations, which can be valuable for clinical decision making and trial design. Moreover, we have identified additional predictors, particularly genetic factors such as GBA and LRRK2 variants, which add new dimensions to our understanding of MF risk. By integrating several of these factors into robust predictive models trained and validated across distinct cohorts, we offer a tool that could improve the design of precision medicine trials and potentially inform personalized patient management strategies. Future research should focus on prospective validation of these models and exploration of their utility in clinical practice and trial design.

Author Roles

Rebecca Ting Jiin Loo: Writing original draft, visualization, validation, methodology, formal analysis. Lukas Pavelka: Data collection, review & editing. Graziella Mangone: review & editing. Fouad Khoury: review & editing. Marie Vidailhet: review & editing. Jean‐Christophe Corvol: review & editing. Rejko Krüger: review & editing. Enrico Glaab: Review & major editing of all parts of the manuscript, supervision, methodology, investigation, funding acquisition.

Ethics Statement

All participants enrolled in the Luxembourg Parkinson's Study provided written informed consent. The study received approval from the National Research Ethics Committee (CNER Ref: 201407/13) and adhered to the principles outlined in the Declaration of Helsinki. Additionally, the Luxembourg Parkinson's Study is registered with ClinicalTrials.gov under the identifier NCT05266872.

Supporting information

Data S1.

Fig. S1. The study workflow includes single‐cohort and multi‐cohort analyses, focusing on motor fluctuations (MF) classification and time‐to‐MF analysis to identify key predictors for MF. In the single‐cohort analysis, predictive models were developed and validated separately for the LuxPARK, PPMI, and ICEBERG cohorts, identifying cohort‐specific predictors. The multi‐cohort analysis integrated data across all cohorts to increase the generalizability and robustness of the predictions.

Fig. S2. Workflow for data processing and model development, illustrating the process, including variable aggregation and the nested cross‐validation workflow (Fig. 1) for model training and evaluation. Data processing steps, including missing values imputation, cross‐study normalization, one‐hot encoding, under sampling, and feature selection, were conducted independently on the training and validation sets to avoid data leakage, as detailed in the Supplementary Fig. S3. The comprehensive model included all features, while the refined model was trained by excluding baseline motor fluctuations and levodopa intake.

Fig. S3. Overview of the data processing and analysis workflow applied during each cross‐validation cycle to optimize and evaluate the prognostic models. The workflow covers multiple steps, including missing value imputation, cross‐study normalization, one‐hot encoding, under sampling, and feature selection.

Fig. S4. A comparison of cross‐validated AUC scores and probabilities for superior predictive performance for the best comprehensive classification model for motor fluctuations in the multi‐cohort analyses. The upper section presents boxplots of cross‐validated AUC scores for each cohort, with the dotted line indicating the hold‐out AUC scores. The lower section shows probabilities of one cohort's predictive performance surpassing another's. The arrows indicate a higher probability of predictive performance, highlighting that the cohort with a higher likelihood of superior performance for the best model is the one towards which the arrows point.

Fig. S5. A comparison of cross‐validated AUC scores and probabilities for superior predictive performance for the best refined classification model for motor fluctuations in the multi‐cohort analyses. The upper section presents boxplots of cross‐validated AUC scores for each cohort, with the dotted line indicating the hold‐out AUC scores. The lower section shows probabilities of one cohort's predictive performance surpassing another's. The arrows indicate a higher probability of predictive performance, highlighting that the cohort with a higher likelihood of superior performance for the best model is the one towards which the arrows point.

Fig. S6. A comparison of cross‐validated C‐indices and probabilities for superior predictive performance for the best comprehensive time‐to‐MF model in the multi‐cohort analyses. The upper section presents boxplots of cross‐validated C‐indices for each cohort, with the dotted line indicating the hold‐out C‐indices. The lower section shows probabilities of one cohort's predictive performance surpassing another's. The arrows indicate a higher probability of predictive performance, highlighting that the cohort with a higher likelihood of superior performance for the best model is the one towards which the arrows point.

Fig. S7. A comparison of cross‐validated C‐indices and probabilities for superior predictive performance for the best refined time‐to‐MF model in the multi‐cohort analyses. The upper section presents boxplots of cross‐validated C‐indices for each cohort, with the dotted line indicating the hold‐out C‐indices. The lower section shows probabilities of one cohort's predictive performance surpassing another's. The arrows indicate a higher probability of predictive performance, highlighting that the cohort with a higher likelihood of superior performance for the best model is the one towards which the arrows point.

Fig. S8. Stability analysis for the comprehensive models for predicting motor fluctuations (MF) in Parkinson's disease (PD) across different algorithms and cohort studies. The stability of the model is evaluated by calculating the standard deviations of the area under the curve (AUC) values across the cross‐validation cycles. A lower standard deviation (SD) indicates a higher stability of the predictive models.

Fig. S9. Stability analysis of comprehensive time to motor fluctuation (MF) models in Parkinson's disease (PD) across different algorithms and cohort studies. The stability of the model is evaluated by calculating the standard deviations of the C‐indices across the cross‐validation cycles. A lower standard deviation (SD) indicates a higher stability of the predictive models.

Fig. S10. Stability analysis of refined models for predicting motor fluctuations (MF) in Parkinson's disease (PD) across different algorithms and cohort studies. The stability of the model is evaluated by calculating the standard deviations of the area under the curve (AUC) values across the cross‐validation cycles. A lower standard deviation (SD) indicates a higher stability of the predictive models.

Fig. S11. Stability analysis of refined time to motor fluctuation (MF) models in Parkinson's disease (PD) across different algorithms and cohort studies. The stability of the model is evaluated by calculating the standard deviations of the C‐indices across the different cross‐validation cycles. A lower standard deviation (SD) indicates a higher stability of the predictive models.

Fig. S12. SHAP value plot illustrating the most informative predictors for the best refined model for cross‐cohort classification of motor fluctuations. SHAP (SHapley Additive exPlanations) plot, providing a detailed view of how individual predictors influence the model's predictions. This visualization allows us to understand not just the overall importance of each predictor, but also how different values of that predictor impact the likelihood of motor fluctuations across our dataset. Each row represents a predictor, ordered by overall importance from top to bottom, and each point represents an individual observation in the dataset. The color of the points ranges from blue (lower values) to red (higher values) for that predictor. The position of points on the x‐axis shows the impact on the model's prediction: Points to the right of zero indicate the predictor pushed the model towards predicting motor fluctuations, and points to the left of zero indicate the predictor pushed the model away from predicting motor fluctuations. Thus, the SHAP plot shows whether high or low values of a predictor are associated with a higher or lower likelihood of predicted motor fluctuations.

Fig. S13. SHAP value plot illustrating the most informative predictors for the best refined model in the cross‐cohort time‐to‐motor fluctuation analysis. SHAP (SHapley Additive exPlanations) plot, providing a detailed view of how individual predictors influence the model's predictions. This visualization allows us to understand not just the overall importance of each predictor, but also how different values of that predictor impact the likelihood of motor fluctuations across our dataset. Each row represents a predictor, ordered by overall importance from top to bottom, and each point represents an individual observation in the dataset. The color of the points ranges from blue (lower values) to red (higher values) for that predictor. The position of points on the x‐axis shows the impact on the model's prediction: Points to the right of zero indicate the predictor pushed the model towards predicting motor fluctuations, and points to the left of zero indicate the predictor pushed the model away from predicting motor fluctuations. Thus, the SHAP plot shows whether high or low values of a predictor are associated with a higher or lower likelihood of predicted motor fluctuations.

Fig. S14. Bar plot illustrating the area under the positive net benefit curve for various best cross‐cohort comprehensive motor fluctuations classification models. The lines indicate significant differences in the net benefit area across the models. The blue bars represent models with a larger positive net benefit area than the negative net benefit, whereas the red bars indicate the opposite. The numbers within the bars represent the difference in net benefit area relative to the “all intervention” strategy. Arrows pointing upwards (↑) indicate a larger area than the “all intervention” strategy, while arrows pointing downwards (↓) indicate a smaller area.

Fig. S15. Bar plot illustrating the area under the positive net benefit curve for various best cross‐cohort refined motor fluctuations classification models. The lines indicate significant differences in the net benefit area across the models. The blue bars represent models with a larger positive net benefit area than the negative net benefit, whereas the red bars indicate the opposite. The numbers within the bars represent the difference in net benefit area relative to the “all intervention” strategy. Arrows pointing upwards (↑) indicate a larger area than the “all intervention” strategy, while arrows pointing downwards (↓) indicate a smaller area.

Fig. S16. Bar plot illustrating the area under the positive net benefit curve for various best cross‐cohort comprehensive time‐to‐MF models. The lines indicate significant differences in the net benefit area across the models. The blue bars represent models with a larger positive net benefit area than the negative net benefit, whereas the red bars indicate the opposite. The numbers within the bars represent the difference in net benefit area relative to the “all intervention” strategy. Arrows pointing upwards (↑) indicate a larger area than the “all intervention” strategy, while arrows pointing downwards (↓) indicate a smaller area.

Fig. S17. Bar plot illustrating the area under the positive net benefit curve for various best cross‐cohort refined time‐to‐MF models. The lines indicate significant differences in the net benefit area across the models. The blue bars represent models with a larger positive net benefit area than the negative net benefit, whereas the red bars indicate the opposite. The numbers within the bars represent the difference in net benefit area relative to the “all intervention” strategy. Arrows pointing upwards (↑) indicate a larger area than the “all intervention” strategy, while arrows pointing downwards (↓) indicate a smaller area.

Table S1. For each considered cohort and their combination (see column 1), the following statistics are provided: Column 2: The numbers of PD patients who met inclusion criterion 1, i.e. a diagnosis of Parkinson's disease according to the UK Parkinson's Disease Society Brain Bank Diagnostic Criteria (UKPDSBB) criteria, or the presence of at least two of the following symptoms: resting tremor, bradykinesia, or rigidity with either resting tremor or bradykinesia, or a single asymmetric rest tremor or asymmetric bradykinesia (see also the Methods sub‐section ‘Study population’ in the main manuscript). Column 3: The number of PD patients who met inclusion inclusion criterion 2, i.e. subjects with the confirmed presence of motor fluctuations within four years following the baseline visit or the completion of a motor fluctuations assessment without presenting symptoms within four years. Columns 3 and 4: The “Events” columns indicate the number and percentage of the patients for whom motor fluctuations were detected within the 4‐year follow‐up period considered for classification analysis (Column 3) and the time‐to‐MF analyses (Column 4) for each cohort separately (first three rows) and across all cohorts combined (fourth row).

Table S2. Overview of the demographic and baseline clinical characteristics for the subjects who developed MF (MF+) or did not develop MF (MF−) in the cross‐cohort analysis during a 4‐year follow‐up period. p‐Values for the significance of differences between the MF− and MF+ groups for individual features are shown in the last column.

Table S3. Overview of the demographic and baseline clinical characteristics for the subjects who developed MF (MF+) or did not develop MF (MF−) in the LuxPARK cohort during a 4‐year follow‐up period. p‐Values for the significance of differences between the MF− and MF+ groups for individual features are shown in the last column.

Table S4. Overview of the demographic and baseline clinical characteristics for the subjects who developed MF (MF+) or did not develop MF (MF−) in the PPMI cohort during a 4‐year follow‐up period. P‐values for the significance of differences between the MF− and MF+ groups for individual features are shown in the last column.

Table S5. Overview of the demographic and baseline clinical characteristics for the subjects who developed MF (MF+) or did not develop MF (MF−) in the ICEBERG cohort during a 4‐year follow‐up period. P‐values for the significance of differences between the MF− and MF+ groups for individual features are shown in the last column.

Table S6. The definition of aggregated feature variables derived from original MDS‐UPDRS variables, where missing values were addressed by averaging non‐missing related variables. Items retained for analysis were marked with an asterisk (*), ensuring that the data were representative and reducing the potential for bias.

Table S7. Summary of predictive performance metrics for comprehensive motor fluctuations classification. The table presents cross‐validated and hold‐out AUC values for single and multi‐cohort analyses and the number of features used in each model. Models with the highest average cross‐validated AUC scores in each cohort analysis across the nested CV hyper parameter optimizations are highlighted in bold and are referred to as the best models. The model with the highest hold‐out AUC score is indicated in italics. The “number of features” column includes the number of candidate features selected during cross‐validation (shown in brackets) and the subset of features demonstrating significant predictive impact identified through permutation importance analysis (preceding the brackets).

Table S8. Summary of predictive performance statistics for refined motor fluctuation classification (excluding baseline motor fluctuations and levodopa medication as input features). The table presents cross‐validated and hold‐out AUC values for single and multi‐cohort analyses and the number of features used in each model. Models with the highest average cross‐validated AUC scores in each cohort analysis are highlighted in bold and are referred to as the best models. The model with the highest hold‐out AUC score is indicated in italics. The “number of features” column includes the number of candidate features selected during cross‐validation (shown in brackets) and the subset of features demonstrating significant predictive impact identified through permutation importance analysis (preceding the brackets).

Table S9. Summary of predictive performance statistics for refined time‐to‐motor fluctuations (excluding baseline motor fluctuations and levodopa medication as input features). The table presents cross‐validated and hold‐out C‐indices for single and multi‐cohort analyses and the number of features used in each model. Models with the highest average cross‐validated C‐indices in each cohort analysis are highlighted in bold and are referred to as the best models. The model with the highest hold‐out C‐index is indicated in italics. The “number of features” column includes the number of candidate features selected during cross‐validation (shown in brackets) and the subset of features demonstrating significant predictive impact identified through permutation importance analysis (preceding the brackets).

Table S10. Comparative analysis of baseline features’ mean differences between the LuxPARK, PPMI, and ICEBERG cohorts. The p‐values highlight statistically significant differences in predictor averages between specific cohort pairs, revealing variations in predictor distributions specific to each cohort in motor fluctuations analysis.

Table S11. Statistics on the average percentage of times predictors were selected during 5‐fold cross‐validation analyses. It compares data for the best comprehensive and refined models in motor fluctuations classification and time‐to‐motor fluctuations analyses across the LuxPARK, PPMI, and ICEBERG cohorts. The information presented includes: “Average in CV (%)” – the average percentage of times each feature was chosen in 5‐fold CV for single‐cohort analyses in LuxPARK, PPMI, and ICEBERG for both motor fluctuations and time‐to‐motor fluctuations analyses; and “Average (%)” – the mean of the “Average in CV (%)” across all cohorts. Features are listed in descending order based on their overall average selection percentages in the best comprehensive and refined models for motor fluctuations and time‐to‐motor fluctuations analyses, with the top 15 most consistent features presented.

Table S12. Statistics on the average percentage of times predictors were selected during 5‐fold cross‐validation analyses across the machine learning algorithms. It compares data for the best comprehensive and refined models for each machine learning algorithm in motor fluctuations classification and time‐to‐motor fluctuations analyses across the LuxPARK, PPMI, and ICEBERG cohorts. The information presented includes: “Average in CV (%)” – the average percentage of times each feature was chosen in 5‐fold CV for single‐cohort analyses in LuxPARK, PPMI, and ICEBERG for both motor fluctuations and time‐to‐motor fluctuations analyses across the algorithms; and “Average (%)” – the mean of the “Average in CV (%)” across all cohorts. Features are listed in descending order based on their overall average selection percentages in the best comprehensive and refined models for motor fluctuations and time‐to‐motor fluctuations analyses, with the top 15 most consistent features presented.

Table S13. Comparison of the statistical significance in hold‐out predictive metrics between the best comprehensive and refined models across cohorts and between the best cross‐study normalized and unnormalized models. The type of normalization applied is indicated in the “Normalization” column. The statistical significance of the differences was assessed using DeLong's test for motor fluctuation classification (top) and a one‐shot nonparametric test for time‐to‐motor fluctuations analysis (bottom). The comprehensive model demonstrated superior predictive performance compared to the refined model, with a statistically significant difference in the hold‐out AUC/C‐index.

Table S14. Evaluation of the predictive performance for comprehensive and refined prognostic models for motor fluctuations (MF), including MF classification and time‐to‐MF analysis models. This evaluation provides cross‐validated and hold‐out AUC values and C‐indices for normalized and non‐normalized models. The “number of features” column indicates both the total number of candidate features selected during cross‐validation (in brackets) and the number of features with significant predictive impact, as determined by permutation importance analysis (preceding the brackets).

Table S15. Top 10 predictors for motor fluctuations (MF) prognosis, ranked by average permutation importance across best comprehensive and refined models for MF classification and time‐to‐MF analysis in the cross‐cohort analysis. The final rank is the average of the non‐missing ranks across the best models.

Table S16. The correlation between predictors and the motor fluctuations occurrence within 4 years and baseline motor fluctuations severity. The correlation is measured using the point‐biserial correlation for continuous or ordinal predictors and Matthews correlation coefficient (MCC) for binary predictors.

Table S17. The results of the correlation analysis of the predictors. The Spearman correlation coefficient was used for continuous or ordinal variable pairs, the point‐biserial correlation coefficient was used for continuous or ordinal and binary variable pairs, and the Matthews correlation coefficient (MCC) was used for binary variable pairs. p‐Values in parentheses accompany the correlation coefficients are presented to indicate the statistical significance of the correlation.

Table S18. Summary statistics for the levodopa equivalent daily dose (LEDD) among MF− and MF+ patients with PD in LuxPARK cohort, along with the statistical significance of the observed differences is indicated by p‐values derived from t‐tests (for normally distributed data) or Mann‐Whitney U‐tests (for data that is not normally distributed). The table also presents time‐to‐MF statistics for PD patients with LEDD < 400 mg and LEDD ≥ 400 mg, including p‐values from log‐rank tests comparing these two groups.

Table S19. The calibration analysis results for comprehensive and refined models in both motor fluctuations (MF) classification and time‐to‐MF analysis in cross‐cohort analysis. The calibration slope and mean squared error (MSE) illustrate the agreement between predicted probabilities and observed outcomes. Slope of 1 indicates perfect calibration.

Data Availability Statement

The LuxPARK clinical dataset used in this study was obtained from the NCER‐PD. The dataset for this manuscript is not publicly available as it is linked to the Luxembourg Parkinson's Study and its internal regulations. Any requests for accessing the dataset can be directed to request.ncer-pd@uni.lu. Data used in the preparation of this article were obtained on January 11, 2023, from the PPMI database (https://www.ppmi-info.org/access-data-specimens/data, RRID:SCR 006431). For up‐to‐date information on the study, please visit the PPMI website (www.ppmi-info.org). Data from the ICEBERG cohort analyzed during this study is available from the corresponding study group (jean-christophe.corvol@aphp.fr, marie.vidailhet@aphp.fr). The data processing, normalization and statistical analyses were performed using the R statistical programming language (v4.2.1). Python‐3.8.6‐GCCcore‐10.2.0 was used for efficient machine learning predictions. The open‐source code is accessible in the GitLab repository under the MIT license: https://gitlab.com/uniluxembourg/lcsb/biomedical-data-science/bds/ml-motor-fluctuations.