Visualization of Decision Tree State for the Classification of Parkinson's Disease

Abstract

Decision trees have been shown to be effective at classifying subjects with Parkinson's disease when provided with features (subject scores) derived from FDG-PET data. Such subject scores have strong discriminative power but are not intuitive to understand. We therefore augment each decision node with thumbnails of the principal component (PC) images from which the subject scores are computed, and also provide labeled scatter plots of the distribution of scores. These plots allow the progress of individual subjects to be traced through the tree and enable the user to focus on complex or unexpected classifications. In addition, we present a visual representation of a typical brain activity pattern arriving at each leaf node, and show how this can be compared to a known reference to validate the behaviour of the tree.

Target and (Astro-)WISE technologies Data federations and its applications

“…The construction of decision trees can be a complex process, however, the resulting trees are typically simple and can be understood by users with little or no background in machine learning. Principal component analysis is combined with the Scaled Subprofile Model to obtain a set of subject scores for each subject which are used as features in decision tree classification (Williams et al 2016). A problem is that the subject scores are not very intuitive.…”