Causal Bayesian network learning: application to the causal analysis of masticatory function and coronary heart...

Coronary heart disease (CHD) is a leading global cause of death, with established links to various risk factors. Recent evidence suggests an association between impaired masticatory function and CHD, but the causal nature of this relationship remains unclear.

To investigate the causal relationship between masticatory function and CHD, we propose a novel hybrid approach that combined constraint-based and score-based methods to learn causal Bayesian network (CBN) structures from observational data. The proposed method is compared against PC, TAN, Naive Bayes, Chow-Liu, and HillClimb algorithms using BIC, K2, BDeu, and BDs scores. A cross-sectional study was conducted in Shanghai, China, from January to May 2023. Data were collected via structured questionnaires, yielding a sample of 179,141 individuals. For analysis, 10,817 healthy individuals (free of any chronic diseases) and 3,382 individuals with CHD are selected. A total of 15 variables covering sociodemographic, functional, cognitive, and health-related aspects are included in the CBN modeling to capture relevant confounders and mediators.

The proposed method outperforms several baseline algorithms while maintaining model interpretability. The learned CBN identifies age, motor function, medical constipation, and sleep as direct causes of CHD. Masticatory function is found to influence CHD indirectly through its effects on motor function and medical constipation. Probabilistic inference indicates that sleep, muscle force, cognitive function and education level are key confounders, with decreasing levels of confounding impact. Causal inference using do-calculus reveals that abnormal masticatory function increases CHD risk by 23% points after adjusting for confounding impact, compared to 39% points without adjustment.

This hybrid CBN approach effectively uncovers interpretable causal pathways, emphasizing the role of oral health in cardiovascular risk and demonstrating the utility of causal models in informing public health strategies.