Max-margin weight learning for medical knowledge network.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Intelligent diagnosis performance relies heavily on medical knowledge application.
• Learning medical knowledge weights is crucial for probabilistic graphical models (PGMs).

Purpose of the Study:

• To investigate a discriminative weight-learning method for medical knowledge networks (MKNs).
• To develop a novel approach for enhancing intelligent medical diagnosis systems.

Main Methods:

• Proposed the maximum margin medical knowledge network (M³KN) training model.
• Transformed weight learning into a margin optimization problem using a reasonable margin definition.
• Employed sequential minimal optimization (SMO) and Markov network properties to solve the optimization problem.

Main Results:

• M³KN achieved higher F-measure scores than maximum likelihood learning on Chinese Electronic Medical Records (CEMRs) and Blood Examination Records (BERs).
• The proposed approach demonstrated superiority over classical machine learning algorithms in medical diagnosis.
• Diagnostic accuracy increased with the number of learned CEMRs, highlighting the importance of domain knowledge.

Conclusions:

• The M³KN method reliably learns medical knowledge weights.
• M³KN outperformed existing methods with F-measures of 0.731 for CEMRs and 0.4538 for BERs.
• M³KN shows potential to advance intelligent healthcare investigations.