Modelling long-term progression of Parkinson’s

Clinical problem

Parkinson’s disease (PD) is a neurodegenerative disorder with a long disease trajectory where the rate of disease progression strongly varies between patients. One factor that may partly explain this variation – and is a suitable target for treatment – is a patient’s cardiovascular risk profile. Previous cross-sectional studies have indeed shown an association between high cardiovascular risk and impaired cognition and balance in patients with PD. However, it remains unclear whether there is a causal effect, and whether clinicians should consider cardiovascular interventions in individual PD patients (e.g. statins, anti-hypertensive medication) to slow down the disease progression.

In this project, we use a unique database consisting of 13 longitudinal PD studies, curated by the Critical Path for Parkinson’s Consortium (CPP). This database currently includes 8105 subjects, mostly early PD patients. In addition, the first data from the ongoing Personalized Parkinson Project at the Radboudumc are available. These datasets contain detailed follow-up measurements from multiple domains, including measures for disease progression (e.g. PD symptom scales, cognitive assessments, quality of life), cardiovascular risk factors (e.g. cholesterol levels, blood sugar levels, blood pressure, BMI, smoking history), and cardiovascular treatments (e.g. statins, anti-hypertensive medication).

Solution

The aim of this project is to develop a decision support system that offers predictions of the causal effect of cardiovascular risk management (prescribing statins and/or anti-hypertensive medication) on long-term progression of Parkinson’s disease symptoms. Predictions will be tailored to a number of clinically recognizable profiles.

In order to use longitudinal observational data to build the underlying model, standard statistical methods are insufficient to address potential confounding. Instead, we will use various advanced causal methods to correct for time-varying confounding. These methods will exploit random variation in the decision to start treatment with statins or anti-hypertensive medication.

Tasks

  • Build a model for the effect of cardiovascular risk management on the progression of PD symptoms, that takes into account time-varying confounding (using causal methods such as inverse probability of treatment weighting, g-formula, and/or causal forests).
  • Work together with clinical experts to make sure relevant confounders are included, and that the different types of measurements are appropriately processed.
  • Combine data from multiple datasets using transfer learning techniques, e.g. deal with comparable but not identical outcome measures, differences in the length of follow-up, timing of study visits, etc.
  • Build a tool to visualize the model’s predictions in a way that is interpretable for clinicians.

Innovation

This project will result in a decision support tool that offers personalized predictions of the effect of cardiovascular risk management on the progression of PD symptoms, based on representative observational data. If clinically relevant effect are found, this tool may be used by neurologists as complimentary source of evidence in the treatment of patients with PD. In addition, it may contribute to the scientific basis for conducting more expensive randomized controlled trials.

Methods

The existing research was critically reviewed and Parkinson's disease progression was modelled with linear mixed models. The causal effect of diabetes, blood pressure, cholesterol, body mass index, cardiovascular risk score and a cardiovascular event in the past was assessed on the PPMI cohort in separate models using DAGs. Progression was measured using motor scores (MDS-UPDRS part III) and cognitive assessments (MoCA, LNS, SDMT) over time.

Results

BMI was the only cardiovascular risk factor that had a causal effect on Parkinson's disease progression (0.057 [0.014 – 0.099] extra points on the MDS-UPDRS part III score per BMI point, per year). These results could be reproduced on the Tracking Parkinson's cohort. The full results can be found in table 4.2 t/m 4.5 of this final report.

Conclusion

Although previous research showed evidence of an effect of increased cardiovascular risk on faster Parkinson's disease progression, this research showed that this effect disappears when confounding effects of for example age are taken into account. Only BMI showed a causal effect on the progression of Parkinson's disease. The estimated effect is 0.057 extra points on the MDS-UPDRS part III score of per BMI point, per year. The difference between someone with a BMI of 20 (normal weight) and someone with a BMI of 28 (overweight) is thus clinically important after 5.5 years (equivalent to 2.5 points) according to the model. This effect of BMI on progression was absent in healthy controls, suggesting a Parkinson's disease specific effect. It is not fully clear via what mechanism(s) this aggravation occurs, both increased inflammation and/or less brain reserve are possibilities. The feasibility of investigating the effect of a subsequent, clearly defined intervention on BMI (deliberate weight loss) using only observational data was also assessed. Even when appropriately accounting for time-varying confounding using G methods, we deemed the causal effect to be difficult to identify using the available datasets. Namely, Parkinson's disease progression is associated with weight loss itself which means that the weight loss we observe (less weight at t+1 than at t) is not necessarily the deliberate weight loss intervention we are interested in. This lack of successfully modelling an explicit intervention makes the implications for policy not fully clear. At the least, BMI seems to be a factor to consider in the disease prognosis.

People

Max Oosterwegel

Max Oosterwegel

Master student

Data Science, Radboud University

Luc Evers

Luc Evers

Scientific Researcher

Diagnostic Image Analysis Group

Marjan Meinders

Marjan Meinders

Senior researcher

Healthcare Improvement Science, Radboudumc

Lieneke van den Heuvel

Lieneke van den Heuvel

Neurologist in training

Neurology, Radboudumc

Bas Bloem

Bas Bloem

Professor

Neurology, Radboudumc

Tom Heskes

Tom Heskes

Professor

Data Science, Radboud University