Incorporating expert knowledge when learning Bayesian network structure: a medical case study.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Causal Inference

Background:

• Bayesian networks (BNs) are increasingly utilized in medical AI applications.
• Current BN development relies on either automated learning or expert input, with limited exploration of hybrid approaches.
• Combining automated causal discovery with expert knowledge offers potential for more robust BN construction.

Purpose of the Study:

• To investigate and present methods for hybrid structure learning in Bayesian networks.
• To develop techniques for assessing the results of combined automated and expert-driven BN learning.
• To integrate hybrid learning into a comprehensive knowledge engineering workflow.

Main Methods:

• Utilized the CaMML causal discovery system with public medical data (IOWA dataset).
• Incorporated varying levels of expert prior knowledge into the automated learning process.
• Developed adjacency matrices for result interpretation and an algorithm for generating a single BN from multiple learned networks.

Main Results:

• A detailed knowledge engineering workflow was established for iterative BN development.
• Simple expert priors, such as partially sorting variables into tiers, proved more effective than no priors or complex priors.
• The proposed method for single BN generation identified relationships missed by other approaches.

Conclusions:

• Hybrid causal learning represents a significant advancement in Bayesian network technology.
• The presented methods facilitate the integration of hybrid learning into knowledge engineering.
• Visualisation and analysis techniques enhance the utility of learned Bayesian networks.