Nonparametric scoring methods as a support decision tool for medical diagnosis – The TreeRank algorithm and its variants

In this paper we propose to use nonparametric scoring methods based on ranking trees as a support decision tool for medical diagnosis. The proposed algorithms enable to order cohorts of patients according to the risk level of developing a particular disease. The aim of this paper is to illustrate the potential of various algorithms using ranking trees, particularly the variants with bagging-type aggregation of these trees, through numerical experiments. The main algorithms presented are: LeafRank (splitting rule), TreeRank (recursive partitioning for ranking) and Ranking Forests (agreggation of ranking trees), as well as the main state-of-the-art methods for bipartite ranking. Numerical experiments based on both artificial and real data sets are provided to discuss the performances of these algorithms in terms of ROC curves and their summaries (AUC or local AUC) but also their properties with respect to interpretability and robustness.