The performance of latent class analysis for clustering multiple long-term conditions is robust to the impact of high-prevalence conditions

Résumé fourni par la source

OBJECTIVES: Living with multiple long-term conditions (MLTC) is increasingly common, posing challenges for health-care systems and individuals alike. Identifying clusters of co-occurring conditions has been proposed as key to understanding disease patterns and supporting patient-centered coordinated care. Latent class analysis (LCA) has been suggested as the optimal clustering method for MLTC, based on simulation studies where condition prevalence and correlations were assumed to be constant across clusters. However, real-world data demonstrate substantial variation in both prevalence and intercondition correlations, particularly when highly prevalent conditions, such as hypertension, coexist with much rarer conditions. The objective of this methodological study was to evaluate the performance and robustness of LCA using real-world data from routinely collected electronic health records (EHRs) for people hospitalized in North-East England. STUDY DESIGN AND SETTING: We investigated the performance and robustness of LCA using information on 60 long-term conditions (LTC). Four analytical approaches were assessed: (1) including all LTC, (2) excluding the most prevalent condition, (3) restricting analyses to the population living with the most prevalent condition and applying LCA to the remaining 59 conditions, and (4) restricting analyses to the population without the most prevalent condition and applying LCA to the remaining 59 conditions. LCA performance was examined using criteria including patient partitioning, condition clustering patterns, and dispersion of condition prevalence across clusters, with further assessment through bootstrap sampling to evaluate reproducibility. RESULTS: Across all approaches, LCA consistently demonstrated strong performance according to these criteria. Excluding or stratifying by the most prevalent condition led to only marginal improvements in the clustering accuracy and stability of the remaining conditions. CONCLUSION: These findings confirm that LCA remains a robust and reliable method for MLTC clustering in realistic settings where prevalence of different LTC varies markedly and a single dominant LTC is observed. This supports the continued use of LCA to understand complex disease patterns and guide future MLTC research. PLAIN LANGUAGE SUMMARY: Many people live with two or more long-term health conditions, which can make their care more complex. Researchers often group patients based on patterns of coexisting conditions to better understand these complexities and improve care planning. One commonly used method for this is called LCA. Previous research has suggested that LCA is a useful way of identifying groups of patients with similar patterns of health conditions. However, this research made assumptions that do not reflect the complexity of real-world health data. In reality, some conditions are very common, while others are rare, and the relationships between different conditions may vary. In this study, we used data from EHRs from a hospital in the UK to test how well LCA performs under more realistic conditions. We also explored whether very common conditions, such as hypertension, affect the results. We found that LCA performs well even when there are large differences in how common conditions are and how they are related to each other. We found no evidence that the presence of one condition, that is much more common than others, influences the results. These findings suggest that LCA is a reliable method for identifying groups of patients with similar patterns of health conditions in real-world data. This can help researchers and clinicians better understand disease patterns and support more personalized and coordinated care for people living with MLTC.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.