Causal-prior-informed prediction of coronary heart disease: Leveraging LLM-guided structure learning and imbalance-aware augmentation

Coronary heart disease (CHD) is a leading cause of death worldwide, making early and accurate detection a critical healthcare priority. However, existing CHD prediction models often face two challenges: severe class imbalance and limited interpretability in characterizing clinically relevant relationships among observed variables. To address these challenges, we propose CausalHeart, a framework that combines MCMC-guided few-shot data augmentation with LLM-informed soft directional priors for graph-conditional CHD risk prediction. The augmentation module improves minority-class representation by selecting representative observed CHD-positive exemplars and curating generated candidate samples. The prediction module incorporates clinically reviewed LLM-derived directional priors as soft regularization terms in directed structure learning and estimates CHD risk conditional on the learned graphical representation. Because the available data are retrospective observational records, the learned structure and associated risk estimates are interpreted as model-based decision-support outputs rather than identified real-world causal effects. Experiments on clinical and external benchmark datasets demonstrate the predictive effectiveness and robustness of the proposed framework under class imbalance. This study contributes a data-efficient and cautiously interpretable approach for CHD risk prediction in high-stakes medical settings.