Use of clustering techniques for clinical and epidemiological research: practical tips using an example from rheumatology

Clinical and epidemiological researchers may want to identify or determine groups with similar characteristics that exist within a dataset or population. For this purpose, clustering techniques can be applied. However, clustering can be complex, and results can be misleading if the methods are not applied with fidelity. The aim of this paper is to provide practical guidance to (clinical) researchers in rheumatology and other fields who wish to explore or confirm (sub) groups within their study population, and are considering to apply clustering techniques to (patient) data. We discuss when clustering is useful for addressing your research question (and when it is not), the need to define the cluster concept, the choice of distance measure between 2 observations, the preprocessing of the data, the assessment of clustering tendency, the types of clustering methods, the determination of the number of clusters and end with the evaluation of the derived clustering solutions using visualizations and statistical measures. The following methods are presented in this paper: K-means, partitioning around medoids, hierarchical clustering, Density-Based Spatial Clustering of Applications with Noise, spectral clustering, fuzzy clustering, and latent class analysis. To illuminate the different considerations involved in clustering, we use a case example: secondary data from 895 people with rheumatoid arthritis, spondyloarthritis, and gout who completed the Health Literacy Questionnaire (HLQ). In this case example, we compute, appraise and evaluate clustering solutions using different methods in 6 steps, including statistical measures and expert opinion. The 6 steps proposed in this paper describe a systematic approach to cluster analysis to ensure all important aspects are considered. In the case example, different clustering methods led to different clustering solutions and insights. Qualitative interpretation by content experts was insightful here. Where possible, researchers should apply more than one clustering method fitting to the objective and reflect on eventual differences in solutions.