Multimorbidity analysis with low condition counts: A robust Bayesian approach for small but important subgroups

Electronic Health Records (EHR) provide a rich source of population data that can be used to systematically understand multimorbidity. Robust measures of association are hypothesis-generating if their co-occurrence is higher than expected. However, prevalent methodological framework are unstable when evidence is limited.

Associations between conditions canbe assembled into a network of multimorbidity, where network analysis can be used to understand the role of diseases, find clusters, or study disease progression. The effects of unreliable association measures get aggregated and intensified when performing network analysis.

These problems affect particularly to minorities (small but important subgroups), for which there is less data available.

We have developed a Bayesian inference framework that is robust to sparse data in the analysis of multimorbidity patterns. The framework also has the following benefits:

- Explicit estimation of uncertainty
- Robust to decisions in the methodology - Robust to noise in rare conditions