Vine Copulas for Analyzing Multivariate Conditional Dependencies in Electronic Health Records Data

TL;DR

This paper introduces a novel application of vine copulas to analyze complex multivariate dependencies in electronic health records, enabling better data exploration and variable selection for healthcare outcomes.

Contribution

It repurposes vine copulas for probabilistic mining of EHR data, providing data-driven explanations, visualization, and variable selection methods.

Findings

Identified conditional dependencies between co-morbid conditions.

Validated the approach on different patient cohorts.

Produced interpretable tree structures representing variable dependencies.

Abstract

Electronic health records (EHR) store hundreds of demographic and laboratory variables from large patient populations. Traditional statistical methods have limited capacity in processing mixed-type data (continuous, ordinal) and capturing non-linear relationships in large multivariate data when oversimplified assumptions are made about the distribution (e.g., Gaussian) of disparate variables in EHR data. This paper addresses the limitations mentioned above by repurposing the vine copula method, which is primarily used to synthesize a multivariate distribution from many bivariate cumulative distribution functions (copulas). Vine copulas produce tree structures that represent bivariate conditional dependencies at varying hierarchical levels, decomposing a multivariate distribution. The tree structure is used to rank variables by conditional dependence and to identify a subset of central variables with local dependence, thus simplifying probabilistic mining of high-dimensional EHR data. The proposed application of vine copulas is used to identify conditional dependence between co-morbid conditions and is validated for characterizing different cohorts of EHR patients. The contribution of this paper is a novel approach to probabilistic mining and exploration of healthcare data that provides data-driven explanations, visualization, and variable selection to prognosticate a healthcare outcome. The source code is shared publicly.