A Bayesian Network Interpretation of the Cox's Proportional Hazard Model

Area of Science:

• Survival Analysis
• Machine Learning
• Biostatistics

Background:

• Cox's proportional hazards (CPH) model is a widely used tool in survival analysis.
• Bayesian networks offer enhanced modeling power and broader applications compared to CPH.
• Often, CPH models exist in literature without access to the original data for re-analysis.

Purpose of the Study:

• To develop a Bayesian network interpretation of the CPH model (BN-Cox).
• To provide a method for encoding knowledge from existing CPH models into Bayesian networks.
• To evaluate the accuracy and efficiency of the BN-Cox model.

Main Methods:

• Encoding knowledge from existing CPH models into Bayesian networks.
• Comparing the accuracy of BN-Cox against CPH, Kaplan-Meier, and other Bayesian networks (Naive Bayes, TAN, Noisy-Max).
• Investigating simplification techniques for BN-Cox: parent divorcing and risk factor removal.

Main Results:

• The BN-Cox model demonstrated high accuracy, closely matching the original CPH model.
• BN-Cox outperformed other Bayesian network approaches in accuracy.
• Simplifying BN-Cox by removing less important risk factors resulted in minimal accuracy loss.

Conclusions:

• BN-Cox provides a viable and accurate Bayesian network alternative for CPH models, especially when original data is unavailable.
• The model offers improved representational and computational efficiency through simplification strategies.
• This approach facilitates knowledge transfer and application of CPH insights within a Bayesian network framework.