Uncovering disease regions using pseudo time-series trajectories on clinical trial data

Abstract

We build pseudo time-series from cross sectional data using a combination of distance metrics, graph theoretical operations and resampling methods. In this paper we explore some extensions of these ideas in order to automatically identify disease regions of interest at key junctions and ‘extreme’ ends of the trajectories. We test these on a number of different medical datasets, in order to explore how applicable the approach is to disease models in general. We focus on two issues in this study: firstly, how to build time-series models from cross-sectional data, and secondly how to automatically identify different disease states along these trajectories, along with the transitions between them. Our results on synthetic data show how the hidden transitional states can indeed be discovered from cross-sectional data and demonstrate the power of the approach on real-world datasets for Glaucoma, Parkinson's Disease and Breast Cancer.