Improving hospital decision making with interpretable associations over datacubes

In this paper we propose a new Classification based on Association Rules (CAR) algorithm that improves the interpretability of the results, works over real data from the electronic health records (EHRs), and allows the study of the patient as a whole. It enables tasks such as the discovery of relationships between diseases, or offering several alternative and reasoned diagnoses for the cases of patients with several diseases that analysed separately could lead to mistaken diagnosis. We aim to achieve several goals: to discover hidden relationships; to improve the interpretability and reduce the complexity of the result; to obtain more reliable diagnosis (getting alternative reasoned diagnoses and higher robustness to noisy rules), and to improve the quality of the classifier avoiding the usual over-fitting problem. To this purpose, we define and exploit hierarchies defined over datacubes dimensions, and change the way the association rules are obtained, and their evaluation at the classification process. To prove the utility of our proposal we have used it in an example of cancer discrimination.