Using Anticipative Hybrid Extreme Rotation Forest to predict emergency service readmission risk

Abstract

This paper provides a real life application of the recently published Anticipative Hybrid Extreme Rotation Forest (AHERF), which is an heterogeneous ensemble classifier that anticipates the correct fraction of instances from each basic classifier architecture to be included in the ensemble. Heterogeneous classifier ensembles aim to profit from the diverse problem domain specificities of each classifier architecture in order to achieve improved generalization over a larger spectrum of problem domains. Given a problem dataset, anticipative determination of the desired ensemble composition is carried out as follows: First, we estimate the performance of each classifier architecture by independent pilot cross-validation experiments on a small subsample of the data. Next, classifier architectures are ranked according to their accuracy results. The likelihood of each classifier architecture instance appearing in the ensemble is computed from this ranking. Finally, while building the ensemble, the architecture of each individual classifier is decided by sampling this likelihood probability distribution. In this paper we provide an application of AHERF to a real life problem. Readmission of patients short time (i.e. 72h) after being released poses a great economical and social challenge, so that many efforts are being addressed to predict and avoid readmission events. We present the results of the application of AHERF over a real life dataset composed of 156,120 admission cases recorded between January 2013 and August 2015. AHERF archives results over or close to 70% sensitivity in the prediction of readmissions for adults and pediatric cases, suggesting that it can be used to build institution specific prediction systems.