This paper presents the first comparison of descriptive segmentation methods for population health management. The aim of descriptive segmentation is to identify heterogeneous segments according to some target observed measure. In healthcare it can be used to understand how utilisation is distributed among a population, and to identify the patient attributes which explain the greatest differences (knowledge of which can help shape segment-tailored services). In reviewing a number of segmentation methods that are both employed on the ground and explored more experimentally within the academic literature, this paper aims to open up a range of options allowing clinicians and managers an informed choice on which approach to use for their situation. Results support the recommendation that decision tree approaches are on-the-whole most suitable, being configurable to local data and providing the best inter-segment discrimination. More basic judgemental splits on patient attributes can be powerful, with the count of chronic conditions being a key variable. Prescribed binning methods such as Bridges to Health are unlikely to achieve high levels of discrimination but do have easily interpretable segments and could be useful for benchmarking. Clustering methods are found to lack discriminative power, which can be attributed to a lack of conceptual appropriateness to the problem.

Population health, Population segmentation, Decision trees, Cluster analysis, Healthcare utilisation