Supervised Clustering for Subgroup Discovery: An Application to COVID-19 Symptomatology

ABSTRACT

Subgroup discovery is a data mining technique that attempts to find interesting relationships between different instances in a dataset with respect to a property of interest. Cluster analysis is a popular method for extracting homogeneous groups from a heterogeneous population, however, it often yields results that are challenging to interpret and action. In this work, we propose a novel, multi-step clustering methodology based on SHAP (SHapley Additive exPlanation) values and dimensionality reduction, for the purpose of subgroup discovery. Our method produces well-separated clusters that can be readily differentiated by simple decision rules, to yield interpretable subgroups in relation to a target variable. We illustrate our approach using self-reported COVID-19 symptom data across 2,479 participants who tested positive for COVID-19, resulting in the identification of 16 distinct symptom presentations. Future work will investigate common demographic and clinical features exhibited by each cluster cohort, and map clusters to outcomes to better understand the clinical presentation, risk factors and prognosis in COVID-19, as a timely and impactful application of this methodology. © 2021, Springer Nature Switzerland AG.