Classification Methods for Ordinal Data and Their Applications in Clinical Research (303265)

Keywords: Ordinal classification, Proportional Odds model, Continuation-ratio model, Constrained Partial-proportional Odds model, Adjacent-category Logistic model, Classification Trees, Support Vector Machines, Neural Network

Ordinal outcomes are common in clinical studies such as disease severity, tumor grade, cancer stage, and so on. Many research questions involve predicting ordinal outcomes. For example, how to use CT imaging features to predict tumor grade. There are several statistical regression models for ordinal data that can be used for both explanatory and predictive purposes: proportional odds model, continuation-ratio model, constrained partial-proportional odds model, and adjacent-category logistic model. In addition, several machine learning methods have been proposed for ordinal responses: classification trees for ordinal response, unimodal neural network, and least square support vector machines based in robust tree decoding. In this presentation, these methods will be introduced and illustrated through several real data examples. How to measure the prediction accuracy for ordinal outcome will also be presented. The performance of these methods will be compared and recommendations for practical applications will be given. R codes for applying these methods are available upon request.