Higher-order disease interactions in multimorbidity measurement: marginal benefit over additive disease summation.

Abstract

Background: Current multimorbidity measures often oversimplify complex disease interactions by assuming a merely additive impact of diseases on health outcomes. This oversimplification neglects clinical observations that certain disease combinations can exhibit synergistic effects. Thus, we aimed to incorporate simultaneous higher-order disease interactions into the validated ICD-coded multimorbidity-weighted index (MICD), to assess for model improvement.

Methods: Health and Retirement Study participants with linked Medicare data contributed ICD-9-CM claims, 1991-2012. Top 20 most prevalent and impactful conditions (based on associations with decline in physical functioning) were assessed through higher order interactions (two-way, three-way). We applied the least absolute shrinkage and selection operator (LASSO) and bootstrapping to identify and retain statistically significant disease interactions. We compared model fit in MICD with and without disease interactions in linear models.

Results: We analyzed 73,830 observations from 18,212 participants (training set N=14,570, testing set N=3,642). MICD without interactions produced an overall R2=0.26. Introducing two-way interactions for the top 10 most prevalent and impactful conditions resulted in a R2=0.27, while expanding to top 20 most prevalent and impactful conditions yielded a R2=0.26. When adding three-way interactions, the same top 10 conditions produced a R2=0.26, while expanding to top 20 conditions resulted in a R2=0.24.

Conclusions: We present novel insights into simultaneous higher-order disease interactions for potential integration into multimorbidity measurement. Incorporating two-way disease interactions for the top 10 most prevalent and impactful conditions showed a minimal improvement in model fit. A more precise multimorbidity index may incorporate both the main effects of diseases and their significant interactions.