A call for caution when using network methods to study multimorbidity: an illustration using data from the Canadian...

Area of Science:

• Gerontology and Public Health
• Computational Epidemiology
• Network Science

Background:

• Multimorbidity, the co-occurrence of multiple chronic diseases, is a growing public health concern, particularly in aging populations.
• Network analysis is increasingly used to understand disease co-occurrence patterns, but methodological choices can influence results.
• The impact of specific analytical decisions on identifying multimorbidity clusters remains underexplored.

Purpose of the Study:

• To investigate how different clustering methods and measures of association affect the identification of multimorbidity clusters.
• To compare network analysis-derived clusters with those identified by clinical experts.

Main Methods:

• Cross-sectional analysis of self-reported data on 24 diseases from 30,097 Canadian adults (aged 45-85).
• Employed 5 clustering methods and 11 association measures for network analysis.
• Utilized the adjusted Rand index (ARI) to quantify cluster similarity and compared results to clinician-defined clusters.

Main Results:

• Significant variability in cluster number (1-24) and disease similarity within clusters was observed across different analytical combinations.
• Adjusted Rand index values ranged from -0.02 to 0.24 when comparing network-derived clusters to clinician-derived clusters, indicating low agreement.
• The choice of clustering algorithm and association measure substantially altered the structure and composition of identified multimorbidity clusters.

Conclusions:

• A systematic evaluation of network analysis methods for binary clustered data, such as diseases, is crucial.
• Disease co-occurrence patterns in older adults may not be consistent, suggesting limitations of generalized clustering approaches.
• The findings underscore the need for individualized care strategies in managing multimorbidity among older adults.