VOLUME 29, ISSUE 4 • DECEMBER 2025.
Interpretable machine learning for cross-cohort prediction of motor fluctuations in Parkinson's disease
Motor fluctuations (MF) are a common and complex complication in Parkinson’s disease (PD), shaped by clinical, genetic, and lifestyle factors. Predicting their onset is particularly challenging because of interindividual variability and systematic differences across patient cohorts. The study “Interpretable Machine Learning for Cross-Cohort Prediction of Motor Fluctuations in Parkinson’s Disease” addresses these challenges by applying interpretable machine learning (ML) techniques to data from three well-characterized PD cohorts (LuxPARK, PPMI, ICEBERG).
A key feature of this work is its cross-cohort design, which evaluates predictors across independent datasets to ensure that results are robust and generalizable. Most prior studies relied on single cohorts with smaller sample sizes, which increases the risk of overfitting and limited generalizability. By contrast, this study integrates multiple cohorts into unified prediction models and applies leave-one-cohort-out validation, providing a stronger foundation for identifying reliable predictors of MF.
The use of interpretable ML models is another new aspect. Rather than relying on uninterpretable “black-box” algorithms, the models highlight how individual variables are associated with MF. To ensure robust and generalizable results across cohorts, multiple ML approaches were applied and compared, including tree-based algorithms for classification and time-to-event analyses by integrating with several cross-cohort normalization approaches.
Through this comparative evaluation, models were identified that achieved reliable MF prediction while offering interpretable and robust predictor rankings, quantified by the frequency of feature selection across cross-validation cycles. A broad set of predictors was examined, including motor and non-motor symptom assessments, clinical features, and genetic factors such as GBA and LRRK2. Consistent feature ranking across cross-validation strengthened confidence that the identified predictors are stable and not cohort-specific artifacts. By comparing multiple algorithms, cross-cohort validation, and emphasis on interpretability, this approach provides a rigorous framework for uncovering key determinants of MF risk in PD and demonstrates how ML can yield actionable, generalizable insights beyond conventional analyses.
One of the most notable findings concerns the commonly used PD medication levodopa. Although levodopa intake has long been considered a key driver of MF, the multivariable cross-cohort models showed that its predictive value was not significant once correlated markers of disease progression such as disease duration, severity, and Hoehn & Yahr (H&Y) stage were accounted for. This indicates that the association between levodopa and MF may not be independent but instead reflects its strong correlation with disease progression. Notably, a recent clinical trial similarly highlighted that MF is closely related to disease progression rather than to levodopa exposure itself. Such insights illustrate the utility of multivariable modeling for revealing complex associations among clinical factors.
In addition to clinical predictors, genetic factors also contributed important insights into MF risk. Cross-cohort analyses revealed that pathogenic GBA mutations were associated with a higher risk of developing MF, reflecting more aggressive disease progression in these carriers. LRRK2 mutations were also linked to MF, though with a smaller hazard ratio. Both GBA and LRRK2 variants have been associated with dyskinesia, a common PD complication related to MF, highlighting the multifaceted impact of genetic variants, disease progression, and symptom severity. These findings emphasize the value of incorporating genetic data into predictive models and demonstrate how cross-cohort analysis can reveal generalizable and clinically meaningful predictors.
Beyond prediction, the models may help to inform clinical trial design and patient management. They could guide risk-based participant selection, refine follow-up schedules, and support early interventions aimed at delaying MF onset. Overall, by integrating diverse predictors into cross-cohort validated models, this study provides a quantitative and generalizable framework for MF prediction in PD that might also serve as a template for the study of other disease outcomes and conditions. 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.
Read more Moving Along: