Simple Explanations to Summarise Subgroup Discovery Outcomes: A Case Study Concerning Patient Phenotyping

Abstract

AbstractPhenotyping is essential in medical research, as it provides a better understanding of healthcare problems owing to the fact that clinical phenotypes identify subsets of patients with common characteristics. Subgroup discovery (SD) appears to be a promising machine learning approach because it provides a framework with which to search for interesting subgroups according to the relations between the individual characteristics and a target value. Each single pattern extracted by SD algorithms is human-readable. However, its complexity (the number of attributes involved) and the high number of subgroups obtained make the overall model difficult to understand. In this work, we propose a method with which to explain SD, designed for the clinical context. We have employed a two-step process in order to obtain SD model-agnostic explanations based on a decision tree surrogate model. The complexity involved in evaluating explainable methods led us to adopt a multiple strategy. We first show how explanations are built, and test a selection of state-of-the-art SD algorithms and gold-standard datasets. We then illustrate the suitability of the method in a clinical use case for an antimicrobial resistance problem. Finally, we study the utility of the method by surveying a small group in order to validate it from a human-centric perspective.KeywordsExplainable artificial intelligenceSubgroup discoveryBiomedical informatics