Interpretable Subgroup Discovery in Sepsis Patient Data Using Beam Search

Abstract

Sepsis is a life-threatening condition that requires timely and accurate diagnosis to improve patient outcomes and resource allocation. While predictive models for sepsis risk exist, they often lack interpretability, limiting their clinical utility. This study investigates the use of subgroup discovery, specifically through a beam search algorithm, to uncover interpretable patterns associated with elevated sepsis risk in emergency department patients. Using a real-world dataset from the St. Antonius Hospital comprising over 27,000 entries and 59 clinical features, beam search was applied across 324 parameter configurations to identify meaningful subgroups. Key parameters such as Beam Width, Depth, Bins, and the StandardQF quality function were varied to explore their impact on subgroup characteristics. Results demonstrate a clear trade-off between subgroup coverage and precision de- pending on the minimum number of required subgroups, with configurations requir- ing more subgroups yielding higher lift values but lower patient coverage. Clinically relevant variables such as CRP, bilirubin, creatinine, and missingness patterns con- sistently emerged in top-performing subgroups. Despite computational limitations that restricted comparisons with exhaustive search methods, this study illustrates the potential of subgroup discovery to provide actionable and interpretable insights for sepsis risk stratification in clinical practice.