Please note the time change to 3pm for the upcoming BRAG only.
The fortnightly BRAG meeting will be held this Thursday the 7th of March at 3pm via Zoom and GP-Y801. This session we will have a presentation by Owen Forbes, who will be talking about his research and experiences in a CSIRO internship. This will include the research itself, as well as setting up the internship and what the experience was like, any recommendations for people considering internships, and how it helped to land a CSIRO postdoc.
Zoom Link: https://qut.zoom.us/j/81579201501?pwd=ZVA1bWk0Y01jb2NkSFlqdG5rUE9MQT09&from=addon
Password: 651747
Owen’s Talk
Title: HDR Internship Experience at CSIRO – Developing Bayesian Trajectory Models for Alzheimer’s Disease Pathology
Abstract: Biomarkers of pathology in Alzheimer’s Disease (AD) are thought to exhibit non-linear sigmoidal trajectories of accumulation of the course of the disease. With AD biomarker data typically characterized by limited timepoints and a relatively short timespan of coverage for each individual, estimating individual non-linear trends proves to be an ill-posed problem. The ‘phase-plane’ method was previously conceived to generate a population-level sigmoid based on individual linear fits. In this research, we developed a Bayesian Hierarchical modelling framework to extend on this work and concurrently model linear fits, estimating a low-order polynomial relationship between means and rates of change, and ensuring uncertainty propagation across different stages of modelling.
By estimating a low-order polynomial relationship between individual intercepts and rates of change, our aim is to develop Bayesian estimation of a population-level sigmoidal trajectory for biomarkers such as amyloid beta (AB) deposition in the brain, measured using positron emission tomography (PET). Our methodology builds on earlier univariate and multivariate frequentist implementations for this problem, with a shift towards the Bayesian hierarchical setting. For the phase-plane fit we use piecewise non-negative polynomials defined to take a value of zero outside of two roots, corresponding to the lower and upper asymptote of the estimated sigmoid. We use this implementation based on its theoretical suitability given prior medical and biology research, and superior performance relative to other functions to fit the phase-plane relationship.
One of the advantages of our Bayesian implementation is the quantification of uncertainty at each step, allowing for estimation of credible intervals for modelled estimates and the propagation of uncertainty through different stages. This approach is especially advantageous in generating probability estimates for clinically relevant inferences, including disease stage assessments and estimated times for disease acceleration based on the properties of the estimated polynomial. We present key results including model comparison across several polynomial orders, comparing the predictive performance of our methods with alternative model configurations and other established methodologies. We also highlight key inferences generated from the best performing model, and demonstrate the utility of our approach for communicating key findings for clinical audiences.
Join Zoom Meeting
https://qut.zoom.us/j/81579201501?pwd=ZVA1bWk0Y01jb2NkSFlqdG5rUE9MQT09&from=addon
Meeting ID: 815 7920 1501
Passcode: 651747
Bayesian Research & Applications Group 