Integrating expert's knowledge constraint of time dependent exposures in structure learning for Bayesian networks

Area of Science:

• Computational Biology
• Statistical Learning
• Causal Inference

Background:

• Learning Bayesian network structures is complex, often requiring expert knowledge integration.
• Dynamic Bayesian networks handle longitudinal data but can misinterpret mediated effects in high-dimensional settings.
• Existing methods may struggle with strong correlations and limited sample sizes, common in fields like oncology.

Purpose of the Study:

• To propose a novel method for Bayesian network structure learning using a priori expert knowledge as hard constraints.
• To address challenges in learning causal graphs with time-dependent exposures and complex variable relationships.
• To improve the accuracy of causal discovery in scenarios with limited data and numerous correlated variables.

Main Methods:

• Developed a structure learning algorithm for Bayesian networks incorporating time-dependent exposure.
• Utilized a priori expert knowledge as hard constraints to guide the graph learning process.
• Compared the proposed method against the state-of-the-art PC-algorithm using simulation studies and a real-world application.

Main Results:

• The proposed method demonstrated superior recovery of true causal graphs compared to the PC-algorithm.
• Integrating hard constraints from expert knowledge significantly improved learning accuracy, even with limited prior information.
• The method showed effectiveness in handling scenarios with strongly correlated variables and time-dependent exposures.

Conclusions:

• A priori expert knowledge, when enforced as hard constraints, is a valuable tool for accurate Bayesian network structure learning.
• The proposed method offers a robust approach for causal discovery in complex, longitudinal datasets, particularly in fields like oncology.
• This work advances the field of causal inference by providing a more reliable method for learning complex network structures.