A Causal Learning-Based Virtual Population Framework for Chronic Disease Analysis

Causal analysis is essential for precise chronic disease management. Chronic diseases are characterized by complex risk-factor interactions, substantial patient heterogeneity, and frequent missingness and measurement uncertainty in real-world medical data. Existing heterogeneity analyses are still largely phenotype-driven, whereas most machine learning methods are correlation-based and lack interpretable mechanism modeling. As a result, mechanism-aware patient stratification remains insufficiently addressed. Methodologically, to uncover more comprehensive causal mechanisms, a causal learning-based virtual population modeling framework is proposed for chronic disease heterogeneity analysis. A biomedical causal prior graph over risk factors is first incorporated to construct patient-level virtual phenotypic profiles, by which each patient is represented relative to a clinically meaningful causal reference. To improve robustness to incomplete and noisy electronic health record data, medically plausible perturbed views are generated through feature masking, random dropout, and mild observation perturbation, and stable patient representations are further learned by contrastive refinement. On this basis, a virtual population is formed in the learned phenotypic space for hidden subgroup discovery. To characterize disease heterogeneity mechanistically, subgroup-specific disease mechanisms are modeled as deviations from a standard pattern, so that common pathophysiological regularities and subgroup-dependent variations can be jointly captured. In the hypertension case study, a new heterogeneity model and novel subtypes are identified, and prognostic relevance, biomarker predictive value, distinct mechanistic patterns, and potential intervention implications are demonstrated. These results suggest that causal virtual population modeling can provide a useful framework for precision chronic disease analysis.