Improving patient clustering by incorporating structured variable label relationships in similarity measures

Area of Science:

• Health Informatics
• Medical Data Analysis
• Computational Biology

Background:

• Patient stratification is crucial for treatment efficacy estimation and patient matching in health research.
• Clinical variables from health records, with structured labels, are used to compute patient similarity.
• The impact of variable label relationships on patient similarity measures is under-explored.

Purpose of the Study:

• To adapt and evaluate weighted Cosine similarity measures that incorporate structured label relationships.
• To compute patient similarities from a medico-administrative database using these enhanced measures.

Main Methods:

• Clustering of patients aged 60 based on annual medicine reimbursements from a French medico-administrative database.
• Comparison of standard Cosine similarity with weighted versions considering variable frequencies and label relationships.
• Patient network construction and clustering using the Markov Cluster algorithm, with evaluation via diagnosis enrichment tests.

Main Results:

• Weighted similarity measures incorporating structured label relationships demonstrated superior performance in identifying similar patients.
• These enhanced measures led to the identification of more clusters associated with diverse diagnostic enrichments.
• Enrichment tests yielded clinically interpretable insights into the identified patient clusters.

Conclusions:

• Considering variable label relationships in patient similarity computations significantly improves patient stratification based on health status.
• This approach offers a more nuanced understanding of patient cohorts for medical research and clinical applications.