Mixed-effects additive Bayesian networks for the assessment of ruptured intracranial aneurysms: Insights from multicenter data.

BACKGROUND

Multicenter clinical datasets are increasingly available but are often analyzed using models that ignore center-specific variability or fail to capture complex dependencies among variables. This is particularly limiting in the context of intracranial aneurysms (IAs), where existing risk scores lack precision and interpretability.

OBJECTIVE

We apply mixed-effects additive Bayesian networks (ABNs) to a large multicenter dataset to model interdependencies among demographic, clinical, and aneurysm-specific features in solitary ruptured IAs, while accounting for heterogeneity across study centers.

METHODS

Data from seven European centers were subsequently harmonized into an observational, cross-sectional cohort of patients with solitary ruptured IAs to mitigate selection bias inherent in cohorts with incidentally discovered unruptured IAs. Mixed-effects ABNs modeled probabilistic dependencies using generalized linear models with random intercepts for study centers. Structure learning was performed via structural Markov chain Monte Carlo sampling with Bayesian information criterion scoring. Model performance was assessed using graphical comparison, intraclass correlation coefficients (ICCs), and predictive metrics.

RESULTS

The mixed-effects ABN produced a more parsimonious network than the pooled model and better captured center-specific variation, particularly for variables with high ICCs (e.g., family history: ICC = 0.369). It also outperformed the pooled ABN in predicting family history (Area Under the Curve = 0.694 vs. 0.585) while yielding clinically interpretable associations, such as the influence of sex, smoking, and IA location on IA size.

CONCLUSION

This application of mixed-effects ABNs reveals that accounting for inter-center heterogeneity is critical for accurately modeling risk factor dependencies in multicenter IA cohorts. This approach yields a more parsimonious network structure by reducing spurious associations found in pooled models. By disentangling patient-level effects from center-specific variations, the model enhances predictive power for heterogeneous variables and provides more reliable, clinically interpretable insights into IA pathophysiology, advancing the potential for personalized risk assessment.