Interpretable Patient Subgrouping Using Trace-Based Clustering

Abstract

Antibiotic resistance in hospitals is a general problem whose solution includes the adaptation of antimicrobial therapies to the local epidemiology. The identification of groups of the population with a common phenotype (by means of their clinical histories) is essential if a hospital is to establish a policy regarding antibiotics. Descriptive pattern mining is effective when carrying out exploratory analyses, such as the design of clustering and subgroup algorithms in order to generate groups of interest. However, the researchers have paid little attention to how these types of algorithms are combined and supervised in medical research. We believe that the implication of clinicians in the process, the interpretability of algorithms and patient traceability of the results obtained are also key requirements. In this work, we propose: (1) to adapt well-known clustering algorithms in order to identify subgroups of patients and (2) a man-in-the-loop methodology so as to carry out this task, thus fulfilling the abovementioned requirements. This proposal is evaluated in the context of a hospital’s antimicrobial stewardship problem.