|
Abstract:
|
A number of hypothetical models have been proposed to describe the progression of biomarkers and symptoms associated with Alzheimer's disease (AD). Current models assume that all AD patients follow a common progression pattern with similar underlying pathology; this is likely an oversimplification. Fitting heterogeneous patient data to a model assuming a single ordering can yield unstable and misleading results. To accurately characterize AD biomarker curves an initial step that groups subjects on a syndrome basis is warranted. We define a syndrome, statistically, as a unique progression pattern of AD biomarkers and propose an unsupervised approach to clustering subjects with pre-symptomatic AD that reflects the multi-dimensionality of their pathology. The technique defines a dissimilarity measure based on where subjects are along the disease process (relative to others) and where their markers are relative to each other (within person), utilizing the time-ordering of events to group subjects. Findings are confirmed via Monte Carlo simulation studies and an application using data from ADNI highlights the need for our novel approach to clustering patients into syndrome groups.
|
