Probabilistic patient risk profiling with pair-copula constructions

Abstract

We propose vine copula-based classifiers for probabilistic risk prediction in perioperative settings. We obtain full joint probability models for mixed continuous-ordinal variables by fitting a separate vine copula to each outcome class, capturing nonlinear and tail-asymmetric dependence. In a cohort of 767 elective bowel surgeries (81 serious vs. 686 non-serious complications), posterior probabilities from the fitted vine classification models are used to allocate patients into low-, moderate-, and high-risk groups. Compared to weighted logistic regression and random forests with stratified sampling, the vine copula-based classifiers achieve up to 10% lower class-specific Brier scores and negative log-likelihoods on the out-of-sample. The vine copula-based classifier identifies a large cohort of true low-risk patients potentially eligible for early discharge. Scenario analyses based on the fitted vine copula models provide interpretable risk profiles, including nonlinear relationships between body mass index, surgery duration, and blood loss, which might remain undetected under linear models. These results demonstrate that vine copula-based classifiers offer a reliable and interpretable framework for individualized, probability-based patient risk profiling. As such, they represent a new, promising tool for data-driven decision-making in perioperative care.