Learning imaging biomarker trajectories from noisy Alzheimer’s disease data using a Bayesian multilevel model

Abstract

Characterising the time course of a disease with a protracted incubation period ultimately requires dense longitudinal studies, which can be prohibitively long and expensive. Considering what can be learned in the absence of such data, we estimate cohort-level biomarker trajectories by fitting cross-sectional data to a differential equation model, then integrating the fit. These fits inform our new stochastic differential equation model for synthesising individual-level biomarker trajectories for prognosis support. Our Bayesian multilevel regression model explicitly includes measurement noise estimation to avoid regression dilution bias. Applicable to any disease, here we perform experiments on Alzheimer’s disease imaging biomarker data — volumes of regions of interest within the brain. We find that Alzheimer’s disease imaging biomarkers are dynamic over timescales from a few years to a few decades.