Parkinson’s disease is a progressive neurodegenerative disorder characterised by a wide range of motor and non-motor features, for which there is no known cause. There is large heterogeneity in both symptom presentation and disease progression which has led to the hypothesis there might be disease subtypes (clusters). In 2018, we published a paper that developed and validated Parkinson’s subtypes on two cohorts using k-means cluster analysis (Lawton, 2018) that found moderate agreement in the actual and predicted clusters. Our subtypes were related to both disease progression and medication response. A recent systematic review identified 25 data driven studies that have tried to determine Parkinson’s subtypes (Mestre, 2021). The 25 published papers show some similarities, however the lack of consistency across publications is a barrier to translating these findings to the clinical setting. Possible reasons behind these differences could be due to (i) analytical methods and/or (ii) data issues. Little research has been published comparing whether different cluster analysis methods applied to the same Parkinson’s cohort will come up with substantively different answers. Data issues include comparing research from different stages of the disease, problems due to selection, lack of harmonisation across measurement instruments and the effects of inter-rater variability.
Aims and Objectives
We aim to answer the following questions:
Analytical issues
- How robust are the findings depending on the different cluster analysis methods?
Data issues
- Can we emulate other published cohorts using subsets of data from two cohorts and does harmonisation of different scales give more similar results?
- Can measurement error explain the moderate level of agreement in our previous work? We will simulate different amounts of inter-rater variability based on our two cohorts to see how this impacts on the observed agreement between subtypes.
Methodology
We will use two large cohorts (Tracking and Discovery) of individuals with early Parkinson’s disease that we already host and have data access permission to explore the aims above. If possible we will try to get access to other published cohorts.
To answer the aims and objectives above the student will use different cluster analysis methods such as hierarchical, non-hierarchical and machine learning methods. The student will become familiar with the creation of prediction models to allow the translation of one subtype definition to another cohort. They will use inter-rater agreement statistics to compare the agreement (using both actual and expected random agreement) between different subtype definitions. The student will also gain experience of simulating data under different assumptions, for example where the misclassification of disease severity is altered by different observers, and the impact of using different methods for scale harmonisation.
Applying the methods above throughout this PhD the student will develop advanced statistical programming skills in a range of packages such as STATA and R. They will become experts at manipulation of data and of applying statistical methods to real-world observational data. The student will gain experience in applying methods and evidence from published journal articles.
Keywords
Parkinson’s disease, cluster analysis, prognostic cohorts, data science
How to apply for this project
This project will be based in Bristol Medical School - Population Health Sciences in the Faculty of Health Sciences at the University of Bristol.
Please visit the Faculty of Health Sciences website for details of how to apply
Continue with Facebook
