Bayesian multilevel modelling for prediction and uncertainty estimation of motor progression in Parkinson’s disease #

Tanja Zerenner, Michael Lawton, Anahita Nodehi, Donald Grosset & the Tracking Parkinson’s team, Michele Hu & the OPDC-Discovery team, Yoav Ben-Shlomo

13:30 Tuesday in 2Q50/51.

Part of the Modelling and inference in public health session.

Abstract #

Quantitative descriptions of Parkinson’s disease trajectories are valuable in providing patients, relatives and caregivers with clear information about expected symptoms, severity, time frame and possible impairments, if desired, and can further serve as a basis for intervention and treatment trials.

Longitudinal cohort studies such as the Oxford Parkinson’s Disease Centre (OPDC) Discovery Cohort, the Tracking Parkinson’s Study or the Parkinson’s Progression Markers Initiative (PPMI) have each recruited hundreds of patients with newly diagnosed Parkinson’s disease and followed them up, collecting a wide range of data including demographic data, environmental and genetic risk factors, motor and non-motor symptoms, as well as imaging, blood samples, etc. with some patients being followed up for over 10 years to date. However, even with such a wealth of data available, quantitative descriptions of the course of PD are - at least at the level of the individual patient - challenging due to the marked heterogeneity of symptoms and disease progression in individual patients. Further, ‘true’ disease trajectories are obscured by short-term fluctuations in patient symptoms, the effect of PD medication on symptoms, differences between clinicians assessing symptoms (‘inter-rater variability’) among other sources of noise and error.

Multi-level models are a popular tool in statistical modeling of disease progression as they allow to account for the non-independence of observations (repeated measures of the same subjects taken at multiple time points) and to describe cohort-level progression (‘fixed effects’) as well as a subject specific derivation thereof (‘random effects’). Utilizing a Bayesian approach to multi-level modeling, we estimate the full, multivariate probability density of model parameters and are hence able to provide complete uncertainty quantification in terms of multivariate probability densities of disease trajectories, both on a cohort level as well as single patient’s level. Such uncertainty quantification helps us to (a) identify relative contributions of different sources of uncertainty and (b) select subgroups for intervention trials based on expected progression rates (e.g., by identifying all patients with a 70% probability of progression with respect to a score of interest being above a chosen threshold).

In the presentation, we discuss common scales and scores describing motor symptoms in PD and cohort- and subject-level progression in relation to these scores, focusing on the respective uncertainty and its visualization and quantification.