A systematic review of multimorbidity clusters in heart failure: Effects of methodologies

Background: Clustering algorithms can identify distinct heart failure (HF) subgroups. The choice of algorithms, modelling process, and input variables can impact clustering outcomes. Therefore, we reviewed analytical methods and variables used in studies that performed clustering in patients with HF.

Methods: We systematically searched CINAHL, COCHRANE, EMBASE, OVID Medline, and Web of Science for eligible articles between inception and April 2023. We included primary studies that identified distinct HF multimorbid subgroups and were appraised for risk-of-bias and against methodological recommendations for cluster analysis. A narrative synthesis was performed.

Results: Our analysis included 43 studies, mostly following a cohort design (n = 34, 79 %) and conducted primarily in Europe (n = 15, 35 %) and North America (n = 13, 30 %). Model-based (n = 22, 48 %), centre-based (n = 10, 22 %), and hierarchical class clustering (n = 9, 20 %) were the most frequently employed algorithms, identifying a range of 2–10 multimorbid clusters. Most studies used a combination of multi-modal parameters (i.e., socio-demographics, biochemistry, clinical characteristics, comorbidities and risk factors, cardiac imaging, and biomarkers) (n = 27, 63 %), followed by disease-based parameters (i.e., comorbidities and risk factors) (n = 11, 26 %) as input variables for clustering. Notably, variables used for clustering reflected cardiovascular and metabolic conditions. The phenogroups identified differed by input variables and algorithms used for clustering. We found substantial quality gaps in developing clustering models, variable selection, reporting of modelling processes, and model validation.

Conclusion: Cluster analysis results differed based on the clustering algorithms used and input variables. This review found substantial gaps in analysis quality and reporting. Implementing a methodological framework to develop, validate, and report clustering analysis can improve the clinical utility and reproducibility of clustering outcomes.