In simulated data and health records, latent class analysis was the optimum multimorbidity clustering algorithm

Linda Nichols; Tom Taverner; Francesca Crowe; Sylvia Richardson; Christopher Yau; Steven Kiddle; Paul Kirk; Jessica Barrett; Krishnarajah Nirantharakumar; Simon Griffin; Duncan Edwards; Tom Marshall

doi:10.1016/j.jclinepi.2022.10.011

In simulated data and health records, latent class analysis was the optimum multimorbidity clustering algorithm

J Clin Epidemiol. 2022 Dec:152:164-175. doi: 10.1016/j.jclinepi.2022.10.011. Epub 2022 Oct 11.

Authors

Affiliations

¹ Research Fellow, Department of Statistics, University of Warwick, Coventry, CV4 7AL, UK.
² Research Fellow, Institute of Applied Health Research, University of Birmingham, B15 2TT, UK.
³ Lecturer in Epidemiology and Health Informatics, Institute of Applied Health Research, University of Birmingham, B15 2TT, UK.
⁴ Emeritus Director, University of Cambridge, Cambridge Biomedical Campus, Cambridge, CB2 0SR, UK.
⁵ Professor of Artificial Intelligence, Nuffield Department of Women's & Reproductive Health, University of Oxford, John Radcliffe Hospital, Oxford, OX3 9DU, UK.
⁶ Director, Health Data Science, AstraZeneca, 1 Francis Crick Avenue, Cambridge, Biomedical Campus, Cambridge, CB2 0AA, UK.
⁷ MRC Investigator, University of Cambridge, Cambridge Biomedical Campus, Cambridge, CB2 0SR, UK.
⁸ Professor of Public Health and Health Informatics, Institute of Applied Health Research, University of Birmingham, B15 2TT, UK.
⁹ Professor of General Practice, Primary Care Unit, Strangeways Research Laboratory Worts Causeway Cambridge CB1 8RN, UK.
¹⁰ Senior Clinical Research Associate, Primary Care Unit, Primary Care Unit, Strangeways Research Laboratory, Worts Causeway, Cambridge, CB1 8RN, UK.
¹¹ Professor of Public Health and Primary Care, Institute of Applied Health Research, University of Birmingham, B15 2TT, UK. Electronic address: T.P.Marshall@bham.ac.uk.

Abstract

Background and objectives: To investigate the reproducibility and validity of latent class analysis (LCA) and hierarchical cluster analysis (HCA), multiple correspondence analysis followed by k-means (MCA-kmeans) and k-means (kmeans) for multimorbidity clustering.

Methods: We first investigated clustering algorithms in simulated datasets with 26 diseases of varying prevalence in predetermined clusters, comparing the derived clusters to known clusters using the adjusted Rand Index (aRI). We then them investigated the medical records of male patients, aged 65 to 84 years from 50 UK general practices, with 49 long-term health conditions. We compared within cluster morbidity profiles using the Pearson correlation coefficient and assessed cluster stability using in 400 bootstrap samples.

Results: In the simulated datasets, the closest agreement (largest aRI) to known clusters was with LCA and then MCA-kmeans algorithms. In the medical records dataset, all four algorithms identified one cluster of 20-25% of the dataset with about 82% of the same patients across all four algorithms. LCA and MCA-kmeans both found a second cluster of 7% of the dataset. Other clusters were found by only one algorithm. LCA and MCA-kmeans clustering gave the most similar partitioning (aRI 0.54).

Conclusion: LCA achieved higher aRI than other clustering algorithms.

Keywords: Clustering methods; Electronic medical records; Hierarchical cluster analysis; K-means; Latent class analysis; Multimorbidity; Multiple correspondence analysis.

MeSH terms

Algorithms*
Cluster Analysis
Humans
Latent Class Analysis
Male
Multimorbidity*
Reproducibility of Results

Abstract

MeSH terms

Grants and funding