Utilization of ordinal response structures in classification with high-dimensional expression data

Molecular diagnosis or prediction of clinical treatment outcome based on high-throughput genomics data is a modern application of machine learning techniques for clinical problems. In practice, clinical parameters, such as patient health status or toxic reaction to therapy, are often measured on an ordinal scale (e.g. good, fair, poor). Commonly, the prediction of ordinal end-points is treated as a multi-class classification problem, disregarding the ordering information contained in the response. This may result in a loss of prediction accuracy. Classical approaches to model ordinal response directly, including for instance the cumulative logit model, are typically not applicable to high-dimensional data. We present hierarchical twoing (hi2), a novel algorithm for classification of high-dimensional data into ordered categories. hi2 combines the power of well-understood binary classification with ordinal response prediction. A comparison of several approaches for ordinal classification on real world data as well as simulated data shows that classification algorithms especially designed to handle ordered categories fail to improve upon state-of-the-art non-ordinal classification algorithms. In general, the classification performance of an algorithm is dominated by its ability to deal with the high-dimensionality of the data. Only hi2 outperforms its competitors in the case of moderate effects. 1998 ACM Subject Classification I.5.2 Design Methodology