Abstract :
[en] Parkinson's disease (PD) is characterized by a wide range of motor and non-motor symptoms, many of which manifest as complications. These include levodopa-induced dyskinesia (LID), motor fluctuations (MF), mild-cognitive impairment in PD (PD-MCI), and patient-reported cognitive decline (PRCD). These complications significantly impact quality of life, contributing to physical disability, reduced independence, increased healthcare costs and caregiver burden. The progression and occurrence of these complications vary considerably among patients. While some PD patients experience these complications early in their disease course, others may not develop them even after prolonged disease duration or exposure to standard therapies like levodopa.
This clinical heterogeneity may result from genetic mutations, demographic characteristics, and individual disease phenotypes, but the underlying mechanisms remain to be explored. Identifying key predictors of these complications is important to understand the factors driving their variability and to develop precision medicine approaches in PD management. Predictive insights can inform therapeutic decision-making, such as adjusting levodopa dosages, optimizing adjunctive therapies, and developing targeted non-pharmacological interventions to mitigate the risk or severity of these complications.
Despite the clinical significance of complications such as LID, MF, PD-MCI, and PRCD, existing predictive models often lack generalizability and robustness, primarily due to biases inherent in single-cohort analyses. These models may not fully capture the complex relationships between clinical variables and PD complications across diverse populations. A multi-cohort analysis addresses these limitations by integrating data from multiple independent studies. This approach increases statistical power with larger sample sizes, reduces cohort-specific biases, and improves model reliability and generalizability. The consistent identification of predictors across diverse populations is a key strength of this analysis, providing more reliable and clinically applicable insights.
This study used machine learning (ML) frameworks to identify potential LID, MF, PD-MCI, and PRCD predictors using data from three independent longitudinal PD cohorts (LuxPARK, PPMI, and ICEBERG). Cross-study normalization was integrated into the ML workflow to enhance the predictive capability and ensure study consistency. This approach mitigates inter-cohort variability, enabling the detection of reliable predictors across diverse cohorts and minimizing the influence of cohort-specific biases.
Incorporating cross-study normalization, interpretable ML, and leave-one-cohort-out validation has enabled the identification of robust and generalizable predictors of these complications. The findings of this study contribute to the understanding of PD complications while providing insights for early detection, risk stratification, and personalized interventions for patients at risk of or experiencing LID, MF, PD-MCI, and PRCD.
Key predictors for PD complications were identified, highlighting distinct and overlapping factors. LID was positively associated with axial symptoms, freezing of gait, and rigidity, while negatively associated predictors included later disease onset, higher body weight, and better visuospatial ability. Predictors of MF included freezing of gait, axial symptoms, and pathogenic GBA and LRRK2 variants, with tremors and later disease onset inversely associated with its development. For the analysis of PD-MCI and PRCD, older age at PD diagnosis, visuospatial deficits, and non-motor symptoms like autonomic dysfunction emerged as significant predictors. Additionally, sex differences were observed in cognitive outcomes, with women displaying better global cognition and less cognitive interference.
Overall, this study offers interpretable ML models for early risk stratification and personalized interventions targeting motor and cognitive complications in PD. Through robust multi-cohort analyses, complications such as LID, MF, PD-MCI, and PRCD can be predicted earlier in the disease course. These findings support the implementation of personalized approaches, including adjusting levodopa dosage according to individual characteristics, optimizing adjuvant therapies, and targeted cognitive interventions for individuals at higher risk. Consequently, these approaches can contribute to enhanced patient outcomes and improved quality of life.
The analysis enables precision medicine in PD management by identifying associations between predictors and these complications. It enables clinicians to stratify patients by risk, design individualized treatment plans, and potentially delay or prevent the onset of complications, thereby preserving quality of life and reducing the burden of advanced disease. Furthermore, integrating predictive models into digital health tools and electronic medical records is a potential benefit, enhancing clinical workflows and decision-making efficiency.
