Decision support from local data: creating adaptive order menus from past clinician behavior

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Clinical Decision Support Systems

Background:

• Guideline-based Clinical Decision Support (CDS) systems face challenges with implementation, outdated information, and inability to address complex care.
• Existing CDS systems often fail to adapt to local clinical practices and patient populations.

Purpose of the Study:

• To develop and evaluate a novel approach using Bayesian Network (BN) learning to generate adaptive, context-specific treatment menus from local order-entry data.
• To create a system that assists in drafting localized CDS content, reducing expert review time.

Main Methods:

• Employed the Greedy Equivalence Search algorithm to learn four domain-specific Bayesian Networks (BNs) from 11,344 patient encounters.
• Developed a system to generate situation-specific, rank-ordered treatment menus from the learned BNs.
• Evaluated the system using hospital-simulation methodology, calculating Area Under the Receiver-Operator Curve (AUC) and average menu position.

Main Results:

• The BN approach generated concise treatment menus (average 3.91-5.83 items).
• Achieved good predictive ability with average AUC ranging from .714 to .844, and above 0.9 for 25% of order types.
• Demonstrated superior performance compared to Association Rule Mining, especially for less frequent orders requiring more context.

Conclusions:

• Local clinical knowledge can be effectively extracted from treatment data using BN learning for enhanced decision support.
• The BN methodology captures complex relationships and provides human-readable networks for expert curation, streamlining CDS content development.
• This approach represents a significant advancement in leveraging local, empirical data to improve healthcare decision support systems.